chore: import upstream snapshot with attribution

tensorflow/tools/android/test/jni/imageutils_jni.cc

@@ -0,0 +1,163 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This file binds the native image utility code to the Java class
+// which exposes them.
+
+#include <jni.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "tensorflow/tools/android/test/jni/rgb2yuv.h"
+#include "tensorflow/tools/android/test/jni/yuv2rgb.h"
+
+#define IMAGEUTILS_METHOD(METHOD_NAME) \
+  Java_org_tensorflow_demo_env_ImageUtils_##METHOD_NAME  // NOLINT
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+JNIEXPORT void JNICALL
+IMAGEUTILS_METHOD(convertYUV420SPToARGB8888)(
+    JNIEnv* env, jclass clazz, jbyteArray input, jintArray output,
+    jint width, jint height, jboolean halfSize);
+
+JNIEXPORT void JNICALL IMAGEUTILS_METHOD(convertYUV420ToARGB8888)(
+    JNIEnv* env, jclass clazz, jbyteArray y, jbyteArray u, jbyteArray v,
+    jintArray output, jint width, jint height, jint y_row_stride,
+    jint uv_row_stride, jint uv_pixel_stride, jboolean halfSize);
+
+JNIEXPORT void JNICALL IMAGEUTILS_METHOD(convertYUV420SPToRGB565)(
+    JNIEnv* env, jclass clazz, jbyteArray input, jbyteArray output, jint width,
+    jint height);
+
+JNIEXPORT void JNICALL
+IMAGEUTILS_METHOD(convertARGB8888ToYUV420SP)(
+    JNIEnv* env, jclass clazz, jintArray input, jbyteArray output,
+    jint width, jint height);
+
+JNIEXPORT void JNICALL
+IMAGEUTILS_METHOD(convertRGB565ToYUV420SP)(
+    JNIEnv* env, jclass clazz, jbyteArray input, jbyteArray output,
+    jint width, jint height);
+
+#ifdef __cplusplus
+}
+#endif
+
+JNIEXPORT void JNICALL
+IMAGEUTILS_METHOD(convertYUV420SPToARGB8888)(
+    JNIEnv* env, jclass clazz, jbyteArray input, jintArray output,
+    jint width, jint height, jboolean halfSize) {
+  jboolean inputCopy = JNI_FALSE;
+  jbyte* const i = env->GetByteArrayElements(input, &inputCopy);
+
+  jboolean outputCopy = JNI_FALSE;
+  jint* const o = env->GetIntArrayElements(output, &outputCopy);
+
+  if (halfSize) {
+    ConvertYUV420SPToARGB8888HalfSize(reinterpret_cast<uint8_t*>(i),
+                                      reinterpret_cast<uint32_t*>(o), width,
+                                      height);
+  } else {
+    ConvertYUV420SPToARGB8888(reinterpret_cast<uint8_t*>(i),
+                              reinterpret_cast<uint8_t*>(i) + width * height,
+                              reinterpret_cast<uint32_t*>(o), width, height);
+  }
+
+  env->ReleaseByteArrayElements(input, i, JNI_ABORT);
+  env->ReleaseIntArrayElements(output, o, 0);
+}
+
+JNIEXPORT void JNICALL IMAGEUTILS_METHOD(convertYUV420ToARGB8888)(
+    JNIEnv* env, jclass clazz, jbyteArray y, jbyteArray u, jbyteArray v,
+    jintArray output, jint width, jint height, jint y_row_stride,
+    jint uv_row_stride, jint uv_pixel_stride, jboolean halfSize) {
+  jboolean inputCopy = JNI_FALSE;
+  jbyte* const y_buff = env->GetByteArrayElements(y, &inputCopy);
+  jboolean outputCopy = JNI_FALSE;
+  jint* const o = env->GetIntArrayElements(output, &outputCopy);
+
+  if (halfSize) {
+    ConvertYUV420SPToARGB8888HalfSize(reinterpret_cast<uint8_t*>(y_buff),
+                                      reinterpret_cast<uint32_t*>(o), width,
+                                      height);
+  } else {
+    jbyte* const u_buff = env->GetByteArrayElements(u, &inputCopy);
+    jbyte* const v_buff = env->GetByteArrayElements(v, &inputCopy);
+
+    ConvertYUV420ToARGB8888(
+        reinterpret_cast<uint8_t*>(y_buff), reinterpret_cast<uint8_t*>(u_buff),
+        reinterpret_cast<uint8_t*>(v_buff), reinterpret_cast<uint32_t*>(o),
+        width, height, y_row_stride, uv_row_stride, uv_pixel_stride);
+
+    env->ReleaseByteArrayElements(u, u_buff, JNI_ABORT);
+    env->ReleaseByteArrayElements(v, v_buff, JNI_ABORT);
+  }
+
+  env->ReleaseByteArrayElements(y, y_buff, JNI_ABORT);
+  env->ReleaseIntArrayElements(output, o, 0);
+}
+
+JNIEXPORT void JNICALL IMAGEUTILS_METHOD(convertYUV420SPToRGB565)(
+    JNIEnv* env, jclass clazz, jbyteArray input, jbyteArray output, jint width,
+    jint height) {
+  jboolean inputCopy = JNI_FALSE;
+  jbyte* const i = env->GetByteArrayElements(input, &inputCopy);
+
+  jboolean outputCopy = JNI_FALSE;
+  jbyte* const o = env->GetByteArrayElements(output, &outputCopy);
+
+  ConvertYUV420SPToRGB565(reinterpret_cast<uint8_t*>(i),
+                          reinterpret_cast<uint16_t*>(o), width, height);
+
+  env->ReleaseByteArrayElements(input, i, JNI_ABORT);
+  env->ReleaseByteArrayElements(output, o, 0);
+}
+
+JNIEXPORT void JNICALL
+IMAGEUTILS_METHOD(convertARGB8888ToYUV420SP)(
+    JNIEnv* env, jclass clazz, jintArray input, jbyteArray output,
+    jint width, jint height) {
+  jboolean inputCopy = JNI_FALSE;
+  jint* const i = env->GetIntArrayElements(input, &inputCopy);
+
+  jboolean outputCopy = JNI_FALSE;
+  jbyte* const o = env->GetByteArrayElements(output, &outputCopy);
+
+  ConvertARGB8888ToYUV420SP(reinterpret_cast<uint32_t*>(i),
+                            reinterpret_cast<uint8_t*>(o), width, height);
+
+  env->ReleaseIntArrayElements(input, i, JNI_ABORT);
+  env->ReleaseByteArrayElements(output, o, 0);
+}
+
+JNIEXPORT void JNICALL
+IMAGEUTILS_METHOD(convertRGB565ToYUV420SP)(
+    JNIEnv* env, jclass clazz, jbyteArray input, jbyteArray output,
+    jint width, jint height) {
+  jboolean inputCopy = JNI_FALSE;
+  jbyte* const i = env->GetByteArrayElements(input, &inputCopy);
+
+  jboolean outputCopy = JNI_FALSE;
+  jbyte* const o = env->GetByteArrayElements(output, &outputCopy);
+
+  ConvertRGB565ToYUV420SP(reinterpret_cast<uint16_t*>(i),
+                          reinterpret_cast<uint8_t*>(o), width, height);
+
+  env->ReleaseByteArrayElements(input, i, JNI_ABORT);
+  env->ReleaseByteArrayElements(output, o, 0);
+}

tensorflow/tools/android/test/jni/object_tracking/config.h

@@ -0,0 +1,300 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_CONFIG_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_CONFIG_H_
+
+#include <math.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+
+namespace tf_tracking {
+
+// Arbitrary keypoint type ids for labeling the origin of tracked keypoints.
+enum KeypointType {
+  KEYPOINT_TYPE_DEFAULT = 0,
+  KEYPOINT_TYPE_FAST = 1,
+  KEYPOINT_TYPE_INTEREST = 2
+};
+
+// Struct that can be used to more richly store the results of a detection
+// than a single number, while still maintaining comparability.
+struct MatchScore {
+  explicit MatchScore(double val) : value(val) {}
+  MatchScore() { value = 0.0; }
+
+  double value;
+
+  MatchScore& operator+(const MatchScore& rhs) {
+    value += rhs.value;
+    return *this;
+  }
+
+  friend std::ostream& operator<<(std::ostream& stream,
+                                  const MatchScore& detection) {
+    stream << detection.value;
+    return stream;
+  }
+};
+inline bool operator< (const MatchScore& cC1, const MatchScore& cC2) {
+    return cC1.value < cC2.value;
+}
+inline bool operator> (const MatchScore& cC1, const MatchScore& cC2) {
+    return cC1.value > cC2.value;
+}
+inline bool operator>= (const MatchScore& cC1, const MatchScore& cC2) {
+    return cC1.value >= cC2.value;
+}
+inline bool operator<= (const MatchScore& cC1, const MatchScore& cC2) {
+    return cC1.value <= cC2.value;
+}
+
+// Fixed seed used for all random number generators.
+static const int kRandomNumberSeed = 11111;
+
+// TODO(andrewharp): Move as many of these settings as possible into a settings
+// object which can be passed in from Java at runtime.
+
+// Whether or not to use ESM instead of LK flow.
+static const bool kUseEsm = false;
+
+// This constant gets added to the diagonal of the Hessian
+// before solving for translation in 2dof ESM.
+// It ensures better behavior especially in the absence of
+// strong texture.
+static const int kEsmRegularizer = 20;
+
+// Do we want to brightness-normalize each keypoint patch when we compute
+// its flow using ESM?
+static const bool kDoBrightnessNormalize = true;
+
+// Whether or not to use fixed-point interpolated pixel lookups in optical flow.
+#define USE_FIXED_POINT_FLOW 1
+
+// Whether to normalize keypoint windows for intensity in LK optical flow.
+// This is a define for now because it helps keep the code streamlined.
+#define NORMALIZE 1
+
+// Number of keypoints to store per frame.
+static const int kMaxKeypoints = 76;
+
+// Keypoint detection.
+static const int kMaxTempKeypoints = 1024;
+
+// Number of floats each keypoint takes up when exporting to an array.
+static const int kKeypointStep = 7;
+
+// Number of frame deltas to keep around in the circular queue.
+static const int kNumFrames = 512;
+
+// Number of iterations to do tracking on each keypoint at each pyramid level.
+static const int kNumIterations = 3;
+
+// The number of bins (on a side) to divide each bin from the previous
+// cache level into.  Higher numbers will decrease performance by increasing
+// cache misses, but mean that cache hits are more locally relevant.
+static const int kCacheBranchFactor = 2;
+
+// Number of levels to put in the cache.
+// Each level of the cache is a square grid of bins, length:
+// branch_factor^(level - 1) on each side.
+//
+// This may be greater than kNumPyramidLevels. Setting it to 0 means no
+// caching is enabled.
+static const int kNumCacheLevels = 3;
+
+// The level at which the cache pyramid gets cut off and replaced by a matrix
+// transform if such a matrix has been provided to the cache.
+static const int kCacheCutoff = 1;
+
+static const int kNumPyramidLevels = 4;
+
+// The minimum number of keypoints needed in an object's area.
+static const int kMaxKeypointsForObject = 16;
+
+// Minimum number of pyramid levels to use after getting cached value.
+// This allows fine-scale adjustment from the cached value, which is taken
+// from the center of the corresponding top cache level box.
+// Can be [0, kNumPyramidLevels).
+static const int kMinNumPyramidLevelsToUseForAdjustment = 1;
+
+// Window size to integrate over to find local image derivative.
+static const int kFlowIntegrationWindowSize = 3;
+
+// Total area of integration windows.
+static const int kFlowArraySize =
+    (2 * kFlowIntegrationWindowSize + 1) * (2 * kFlowIntegrationWindowSize + 1);
+
+// Error that's considered good enough to early abort tracking.
+static const float kTrackingAbortThreshold = 0.03f;
+
+// Maximum number of deviations a keypoint-correspondence delta can be from the
+// weighted average before being thrown out for region-based queries.
+static const float kNumDeviations = 2.0f;
+
+// The length of the allowed delta between the forward and the backward
+// flow deltas in terms of the length of the forward flow vector.
+static const float kMaxForwardBackwardErrorAllowed = 0.5f;
+
+// Threshold for pixels to be considered different.
+static const int kFastDiffAmount = 10;
+
+// How far from edge of frame to stop looking for FAST keypoints.
+static const int kFastBorderBuffer = 10;
+
+// Determines if non-detected arbitrary keypoints should be added to regions.
+// This will help if no keypoints have been detected in the region yet.
+static const bool kAddArbitraryKeypoints = true;
+
+// How many arbitrary keypoints to add along each axis as candidates for each
+// region?
+static const int kNumToAddAsCandidates = 1;
+
+// In terms of region dimensions, how closely can we place keypoints
+// next to each other?
+static const float kClosestPercent = 0.6f;
+
+// How many FAST qualifying pixels must be connected to a pixel for it to be
+// considered a candidate keypoint for Harris filtering.
+static const int kMinNumConnectedForFastKeypoint = 8;
+
+// Size of the window to integrate over for Harris filtering.
+// Compare to kFlowIntegrationWindowSize.
+static const int kHarrisWindowSize = 2;
+
+
+// DETECTOR PARAMETERS
+
+// Before relocalizing, make sure the new proposed position is better than
+// the existing position by a small amount to prevent thrashing.
+static const MatchScore kMatchScoreBuffer(0.01f);
+
+// Minimum score a tracked object can have and still be considered a match.
+// TODO(andrewharp): Make this a per detector thing.
+static const MatchScore kMinimumMatchScore(0.5f);
+
+static const float kMinimumCorrelationForTracking = 0.4f;
+
+static const MatchScore kMatchScoreForImmediateTermination(0.0f);
+
+// Run the detector every N frames.
+static const int kDetectEveryNFrames = 4;
+
+// How many features does each feature_set contain?
+static const int kFeaturesPerFeatureSet = 10;
+
+// The number of FeatureSets managed by the object detector.
+// More FeatureSets can increase recall at the cost of performance.
+static const int kNumFeatureSets = 7;
+
+// How many FeatureSets must respond affirmatively for a candidate descriptor
+// and position to be given more thorough attention?
+static const int kNumFeatureSetsForCandidate = 2;
+
+// How large the thumbnails used for correlation validation are.  Used for both
+// width and height.
+static const int kNormalizedThumbnailSize = 11;
+
+// The area of intersection divided by union for the bounding boxes that tells
+// if this tracking has slipped enough to invalidate all unlocked examples.
+static const float kPositionOverlapThreshold = 0.6f;
+
+// The number of detection failures allowed before an object goes invisible.
+// Tracking will still occur, so if it is actually still being tracked and
+// comes back into a detectable position, it's likely to be found.
+static const int kMaxNumDetectionFailures = 4;
+
+
+// Minimum square size to scan with sliding window.
+static const float kScanMinSquareSize = 16.0f;
+
+// Minimum square size to scan with sliding window.
+static const float kScanMaxSquareSize = 64.0f;
+
+// Scale difference for consecutive scans of the sliding window.
+static const float kScanScaleFactor = sqrtf(2.0f);
+
+// Step size for sliding window.
+static const int kScanStepSize = 10;
+
+
+// How tightly to pack the descriptor boxes for confirmed exemplars.
+static const float kLockedScaleFactor = 1 / sqrtf(2.0f);
+
+// How tightly to pack the descriptor boxes for unconfirmed exemplars.
+static const float kUnlockedScaleFactor = 1 / 2.0f;
+
+// How tightly the boxes to scan centered at the last known position will be
+// packed.
+static const float kLastKnownPositionScaleFactor = 1.0f / sqrtf(2.0f);
+
+// The bounds on how close a new object example must be to existing object
+// examples for detection to be valid.
+static const float kMinCorrelationForNewExample = 0.75f;
+static const float kMaxCorrelationForNewExample = 0.99f;
+
+
+// The number of safe tries an exemplar has after being created before
+// missed detections count against it.
+static const int kFreeTries = 5;
+
+// A false positive is worth this many missed detections.
+static const int kFalsePositivePenalty = 5;
+
+struct ObjectDetectorConfig {
+  const Size image_size;
+
+  explicit ObjectDetectorConfig(const Size& image_size)
+      : image_size(image_size) {}
+  virtual ~ObjectDetectorConfig() = default;
+};
+
+struct KeypointDetectorConfig {
+  const Size image_size;
+
+  bool detect_skin;
+
+  explicit KeypointDetectorConfig(const Size& image_size)
+      : image_size(image_size),
+        detect_skin(false) {}
+};
+
+
+struct OpticalFlowConfig {
+  const Size image_size;
+
+  explicit OpticalFlowConfig(const Size& image_size)
+      : image_size(image_size) {}
+};
+
+struct TrackerConfig {
+  const Size image_size;
+  KeypointDetectorConfig keypoint_detector_config;
+  OpticalFlowConfig flow_config;
+  bool always_track;
+
+  float object_box_scale_factor_for_features;
+
+  explicit TrackerConfig(const Size& image_size)
+      : image_size(image_size),
+        keypoint_detector_config(image_size),
+        flow_config(image_size),
+        always_track(false),
+        object_box_scale_factor_for_features(1.0f) {}
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_CONFIG_H_

tensorflow/tools/android/test/jni/object_tracking/flow_cache.h

@@ -0,0 +1,306 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_FLOW_CACHE_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_FLOW_CACHE_H_
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/optical_flow.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+// Class that helps OpticalFlow to speed up flow computation
+// by caching coarse-grained flow.
+class FlowCache {
+ public:
+  explicit FlowCache(const OpticalFlowConfig* const config)
+      : config_(config),
+        image_size_(config->image_size),
+        optical_flow_(config),
+        fullframe_matrix_(NULL) {
+    for (int i = 0; i < kNumCacheLevels; ++i) {
+      const int curr_dims = BlockDimForCacheLevel(i);
+      has_cache_[i] = new Image<bool>(curr_dims, curr_dims);
+      displacements_[i] = new Image<Point2f>(curr_dims, curr_dims);
+    }
+  }
+
+  ~FlowCache() {
+    for (int i = 0; i < kNumCacheLevels; ++i) {
+      SAFE_DELETE(has_cache_[i]);
+      SAFE_DELETE(displacements_[i]);
+    }
+    delete[](fullframe_matrix_);
+    fullframe_matrix_ = NULL;
+  }
+
+  void NextFrame(ImageData* const new_frame,
+                 const float* const align_matrix23) {
+    ClearCache();
+    SetFullframeAlignmentMatrix(align_matrix23);
+    optical_flow_.NextFrame(new_frame);
+  }
+
+  void ClearCache() {
+    for (int i = 0; i < kNumCacheLevels; ++i) {
+      has_cache_[i]->Clear(false);
+    }
+    delete[](fullframe_matrix_);
+    fullframe_matrix_ = NULL;
+  }
+
+  // Finds the flow at a point, using the cache for performance.
+  bool FindFlowAtPoint(const float u_x, const float u_y,
+                       float* const flow_x, float* const flow_y) const {
+    // Get the best guess from the cache.
+    const Point2f guess_from_cache = LookupGuess(u_x, u_y);
+
+    *flow_x = guess_from_cache.x;
+    *flow_y = guess_from_cache.y;
+
+    // Now refine the guess using the image pyramid.
+    for (int pyramid_level = kMinNumPyramidLevelsToUseForAdjustment - 1;
+        pyramid_level >= 0; --pyramid_level) {
+      if (!optical_flow_.FindFlowAtPointSingleLevel(
+          pyramid_level, u_x, u_y, false, flow_x, flow_y)) {
+        return false;
+      }
+    }
+
+    return true;
+  }
+
+  // Determines the displacement of a point, and uses that to calculate a new
+  // position.
+  // Returns true iff the displacement determination worked and the new position
+  // is in the image.
+  bool FindNewPositionOfPoint(const float u_x, const float u_y,
+                              float* final_x, float* final_y) const {
+    float flow_x;
+    float flow_y;
+    if (!FindFlowAtPoint(u_x, u_y, &flow_x, &flow_y)) {
+      return false;
+    }
+
+    // Add in the displacement to get the final position.
+    *final_x = u_x + flow_x;
+    *final_y = u_y + flow_y;
+
+    // Assign the best guess, if we're still in the image.
+    if (InRange(*final_x, 0.0f, static_cast<float>(image_size_.width) - 1) &&
+        InRange(*final_y, 0.0f, static_cast<float>(image_size_.height) - 1)) {
+      return true;
+    } else {
+      return false;
+    }
+  }
+
+  // Comparison function for qsort.
+  static int Compare(const void* a, const void* b) {
+    return *reinterpret_cast<const float*>(a) -
+           *reinterpret_cast<const float*>(b);
+  }
+
+  // Returns the median flow within the given bounding box as determined
+  // by a grid_width x grid_height grid.
+  Point2f GetMedianFlow(const BoundingBox& bounding_box,
+                        const bool filter_by_fb_error,
+                        const int grid_width,
+                        const int grid_height) const {
+    const int kMaxPoints = 100;
+    SCHECK(grid_width * grid_height <= kMaxPoints,
+          "Too many points for Median flow!");
+
+    const BoundingBox valid_box = bounding_box.Intersect(
+        BoundingBox(0, 0, image_size_.width - 1, image_size_.height - 1));
+
+    if (valid_box.GetArea() <= 0.0f) {
+      return Point2f(0, 0);
+    }
+
+    float x_deltas[kMaxPoints];
+    float y_deltas[kMaxPoints];
+
+    int curr_offset = 0;
+    for (int i = 0; i < grid_width; ++i) {
+      for (int j = 0; j < grid_height; ++j) {
+        const float x_in = valid_box.left_ +
+            (valid_box.GetWidth() * i) / (grid_width - 1);
+
+        const float y_in = valid_box.top_ +
+            (valid_box.GetHeight() * j) / (grid_height - 1);
+
+        float curr_flow_x;
+        float curr_flow_y;
+        const bool success = FindNewPositionOfPoint(x_in, y_in,
+                                                    &curr_flow_x, &curr_flow_y);
+
+        if (success) {
+          x_deltas[curr_offset] = curr_flow_x;
+          y_deltas[curr_offset] = curr_flow_y;
+          ++curr_offset;
+        } else {
+          LOGW("Tracking failure!");
+        }
+      }
+    }
+
+    if (curr_offset > 0) {
+      qsort(x_deltas, curr_offset, sizeof(*x_deltas), Compare);
+      qsort(y_deltas, curr_offset, sizeof(*y_deltas), Compare);
+
+      return Point2f(x_deltas[curr_offset / 2], y_deltas[curr_offset / 2]);
+    }
+
+    LOGW("No points were valid!");
+    return Point2f(0, 0);
+  }
+
+  void SetFullframeAlignmentMatrix(const float* const align_matrix23) {
+    if (align_matrix23 != NULL) {
+      if (fullframe_matrix_ == NULL) {
+        fullframe_matrix_ = new float[6];
+      }
+
+      memcpy(fullframe_matrix_, align_matrix23,
+             6 * sizeof(fullframe_matrix_[0]));
+    }
+  }
+
+ private:
+  Point2f LookupGuessFromLevel(
+      const int cache_level, const float x, const float y) const {
+    // LOGE("Looking up guess at %5.2f %5.2f for level %d.", x, y, cache_level);
+
+    // Cutoff at the target level and use the matrix transform instead.
+    if (fullframe_matrix_ != NULL && cache_level == kCacheCutoff) {
+      const float xnew = x * fullframe_matrix_[0] +
+                         y * fullframe_matrix_[1] +
+                             fullframe_matrix_[2];
+      const float ynew = x * fullframe_matrix_[3] +
+                         y * fullframe_matrix_[4] +
+                             fullframe_matrix_[5];
+
+      return Point2f(xnew - x, ynew - y);
+    }
+
+    const int level_dim = BlockDimForCacheLevel(cache_level);
+    const int pixels_per_cache_block_x =
+        (image_size_.width + level_dim - 1) / level_dim;
+    const int pixels_per_cache_block_y =
+        (image_size_.height + level_dim - 1) / level_dim;
+    const int index_x = x / pixels_per_cache_block_x;
+    const int index_y = y / pixels_per_cache_block_y;
+
+    Point2f displacement;
+    if (!(*has_cache_[cache_level])[index_y][index_x]) {
+      (*has_cache_[cache_level])[index_y][index_x] = true;
+
+      // Get the lower cache level's best guess, if it exists.
+      displacement = cache_level >= kNumCacheLevels - 1 ?
+          Point2f(0, 0) : LookupGuessFromLevel(cache_level + 1, x, y);
+      // LOGI("Best guess at cache level %d is %5.2f, %5.2f.", cache_level,
+      //      best_guess.x, best_guess.y);
+
+      // Find the center of the block.
+      const float center_x = (index_x + 0.5f) * pixels_per_cache_block_x;
+      const float center_y = (index_y + 0.5f) * pixels_per_cache_block_y;
+      const int pyramid_level = PyramidLevelForCacheLevel(cache_level);
+
+      // LOGI("cache level %d: [%d, %d (%5.2f / %d, %5.2f / %d)] "
+      //      "Querying %5.2f, %5.2f at pyramid level %d, ",
+      //      cache_level, index_x, index_y,
+      //      x, pixels_per_cache_block_x, y, pixels_per_cache_block_y,
+      //      center_x, center_y, pyramid_level);
+
+      // TODO(andrewharp): Turn on FB error filtering.
+      const bool success = optical_flow_.FindFlowAtPointSingleLevel(
+          pyramid_level, center_x, center_y, false,
+          &displacement.x, &displacement.y);
+
+      if (!success) {
+        LOGV("Computation of cached value failed for level %d!", cache_level);
+      }
+
+      // Store the value for later use.
+      (*displacements_[cache_level])[index_y][index_x] = displacement;
+    } else {
+      displacement = (*displacements_[cache_level])[index_y][index_x];
+    }
+
+    // LOGI("Returning %5.2f, %5.2f for level %d",
+    //      displacement.x, displacement.y, cache_level);
+    return displacement;
+  }
+
+  Point2f LookupGuess(const float x, const float y) const {
+    if (x < 0 || x >= image_size_.width || y < 0 || y >= image_size_.height) {
+      return Point2f(0, 0);
+    }
+
+    // LOGI("Looking up guess at %5.2f %5.2f.", x, y);
+    if (kNumCacheLevels > 0) {
+      return LookupGuessFromLevel(0, x, y);
+    } else {
+      return Point2f(0, 0);
+    }
+  }
+
+  // Returns the number of cache bins in each dimension for a given level
+  // of the cache.
+  int BlockDimForCacheLevel(const int cache_level) const {
+    // The highest (coarsest) cache level has a block dim of kCacheBranchFactor,
+    // thus if there are 4 cache levels, requesting level 3 (0-based) should
+    // return kCacheBranchFactor, level 2 should return kCacheBranchFactor^2,
+    // and so on.
+    int block_dim = kNumCacheLevels;
+    for (int curr_level = kNumCacheLevels - 1; curr_level > cache_level;
+        --curr_level) {
+      block_dim *= kCacheBranchFactor;
+    }
+    return block_dim;
+  }
+
+  // Returns the level of the image pyramid that a given cache level maps to.
+  int PyramidLevelForCacheLevel(const int cache_level) const {
+    // Higher cache and pyramid levels have smaller dimensions. The highest
+    // cache level should refer to the highest image pyramid level. The
+    // lower, finer image pyramid levels are uncached (assuming
+    // kNumCacheLevels < kNumPyramidLevels).
+    return cache_level + (kNumPyramidLevels - kNumCacheLevels);
+  }
+
+  const OpticalFlowConfig* const config_;
+
+  const Size image_size_;
+  OpticalFlow optical_flow_;
+
+  float* fullframe_matrix_;
+
+  // Whether this value is currently present in the cache.
+  Image<bool>* has_cache_[kNumCacheLevels];
+
+  // The cached displacement values.
+  Image<Point2f>* displacements_[kNumCacheLevels];
+
+  FlowCache(const FlowCache&) = delete;
+  void operator=(const FlowCache&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_FLOW_CACHE_H_

tensorflow/tools/android/test/jni/object_tracking/frame_pair.cc

@@ -0,0 +1,309 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/tools/android/test/jni/object_tracking/frame_pair.h"
+
+#include <float.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+
+namespace tf_tracking {
+
+void FramePair::Init(const int64_t start_time, const int64_t end_time) {
+  start_time_ = start_time;
+  end_time_ = end_time;
+  memset(optical_flow_found_keypoint_, false,
+         sizeof(*optical_flow_found_keypoint_) * kMaxKeypoints);
+  number_of_keypoints_ = 0;
+}
+
+void FramePair::AdjustBox(const BoundingBox box,
+                          float* const translation_x,
+                          float* const translation_y,
+                          float* const scale_x,
+                          float* const scale_y) const {
+  static float weights[kMaxKeypoints];
+  static Point2f deltas[kMaxKeypoints];
+  memset(weights, 0.0f, sizeof(*weights) * kMaxKeypoints);
+
+  BoundingBox resized_box(box);
+  resized_box.Scale(0.4f, 0.4f);
+  FillWeights(resized_box, weights);
+  FillTranslations(deltas);
+
+  const Point2f translation = GetWeightedMedian(weights, deltas);
+
+  *translation_x = translation.x;
+  *translation_y = translation.y;
+
+  const Point2f old_center = box.GetCenter();
+  const int good_scale_points =
+      FillScales(old_center, translation, weights, deltas);
+
+  // Default scale factor is 1 for x and y.
+  *scale_x = 1.0f;
+  *scale_y = 1.0f;
+
+  // The assumption is that all deltas that make it to this stage with a
+  // corresponding optical_flow_found_keypoint_[i] == true are not in
+  // themselves degenerate.
+  //
+  // The degeneracy with scale arose because if the points are too close to the
+  // center of the objects, the scale ratio determination might be incalculable.
+  //
+  // The check for kMinNumInRange is not a degeneracy check, but merely an
+  // attempt to ensure some sort of stability. The actual degeneracy check is in
+  // the comparison to EPSILON in FillScales (which I've updated to return the
+  // number good remaining as well).
+  static const int kMinNumInRange = 5;
+  if (good_scale_points >= kMinNumInRange) {
+    const float scale_factor = GetWeightedMedianScale(weights, deltas);
+
+    if (scale_factor > 0.0f) {
+      *scale_x = scale_factor;
+      *scale_y = scale_factor;
+    }
+  }
+}
+
+int FramePair::FillWeights(const BoundingBox& box,
+                           float* const weights) const {
+  // Compute the max score.
+  float max_score = -FLT_MAX;
+  float min_score = FLT_MAX;
+  for (int i = 0; i < kMaxKeypoints; ++i) {
+    if (optical_flow_found_keypoint_[i]) {
+      max_score = MAX(max_score, frame1_keypoints_[i].score_);
+      min_score = MIN(min_score, frame1_keypoints_[i].score_);
+    }
+  }
+
+  int num_in_range = 0;
+  for (int i = 0; i < kMaxKeypoints; ++i) {
+    if (!optical_flow_found_keypoint_[i]) {
+      weights[i] = 0.0f;
+      continue;
+    }
+
+    const bool in_box = box.Contains(frame1_keypoints_[i].pos_);
+    if (in_box) {
+      ++num_in_range;
+    }
+
+    // The weighting based off distance.  Anything within the bounding box
+    // has a weight of 1, and everything outside of that is within the range
+    // [0, kOutOfBoxMultiplier), falling off with the squared distance ratio.
+    float distance_score = 1.0f;
+    if (!in_box) {
+      const Point2f initial = box.GetCenter();
+      const float sq_x_dist =
+          Square(initial.x - frame1_keypoints_[i].pos_.x);
+      const float sq_y_dist =
+          Square(initial.y - frame1_keypoints_[i].pos_.y);
+      const float squared_half_width = Square(box.GetWidth() / 2.0f);
+      const float squared_half_height = Square(box.GetHeight() / 2.0f);
+
+      static const float kOutOfBoxMultiplier = 0.5f;
+      distance_score = kOutOfBoxMultiplier *
+          MIN(squared_half_height / sq_y_dist, squared_half_width / sq_x_dist);
+    }
+
+    // The weighting based on relative score strength. kBaseScore - 1.0f.
+    float intrinsic_score =  1.0f;
+    if (max_score > min_score) {
+      static const float kBaseScore = 0.5f;
+      intrinsic_score = ((frame1_keypoints_[i].score_ - min_score) /
+         (max_score - min_score)) * (1.0f - kBaseScore) + kBaseScore;
+    }
+
+    // The final score will be in the range [0, 1].
+    weights[i] = distance_score * intrinsic_score;
+  }
+
+  return num_in_range;
+}
+
+void FramePair::FillTranslations(Point2f* const translations) const {
+  for (int i = 0; i < kMaxKeypoints; ++i) {
+    if (!optical_flow_found_keypoint_[i]) {
+      continue;
+    }
+    translations[i].x =
+        frame2_keypoints_[i].pos_.x - frame1_keypoints_[i].pos_.x;
+    translations[i].y =
+        frame2_keypoints_[i].pos_.y - frame1_keypoints_[i].pos_.y;
+  }
+}
+
+int FramePair::FillScales(const Point2f& old_center,
+                          const Point2f& translation,
+                          float* const weights,
+                          Point2f* const scales) const {
+  int num_good = 0;
+  for (int i = 0; i < kMaxKeypoints; ++i) {
+    if (!optical_flow_found_keypoint_[i]) {
+      continue;
+    }
+
+    const Keypoint keypoint1 = frame1_keypoints_[i];
+    const Keypoint keypoint2 = frame2_keypoints_[i];
+
+    const float dist1_x = keypoint1.pos_.x - old_center.x;
+    const float dist1_y = keypoint1.pos_.y - old_center.y;
+
+    const float dist2_x = (keypoint2.pos_.x - translation.x) - old_center.x;
+    const float dist2_y = (keypoint2.pos_.y - translation.y) - old_center.y;
+
+    // Make sure that the scale makes sense; points too close to the center
+    // will result in either NaNs or infinite results for scale due to
+    // limited tracking and floating point resolution.
+    // Also check that the parity of the points is the same with respect to
+    // x and y, as we can't really make sense of data that has flipped.
+    if (((dist2_x > EPSILON && dist1_x > EPSILON) ||
+         (dist2_x < -EPSILON && dist1_x < -EPSILON)) &&
+         ((dist2_y > EPSILON && dist1_y > EPSILON) ||
+          (dist2_y < -EPSILON && dist1_y < -EPSILON))) {
+      scales[i].x = dist2_x / dist1_x;
+      scales[i].y = dist2_y / dist1_y;
+      ++num_good;
+    } else {
+      weights[i] = 0.0f;
+      scales[i].x = 1.0f;
+      scales[i].y = 1.0f;
+    }
+  }
+  return num_good;
+}
+
+struct WeightedDelta {
+  float weight;
+  float delta;
+};
+
+// Sort by delta, not by weight.
+inline int WeightedDeltaCompare(const void* const a, const void* const b) {
+  return (reinterpret_cast<const WeightedDelta*>(a)->delta -
+          reinterpret_cast<const WeightedDelta*>(b)->delta) <= 0 ? 1 : -1;
+}
+
+// Returns the median delta from a sorted set of weighted deltas.
+static float GetMedian(const int num_items,
+                       const WeightedDelta* const weighted_deltas,
+                       const float sum) {
+  if (num_items == 0 || sum < EPSILON) {
+    return 0.0f;
+  }
+
+  float current_weight = 0.0f;
+  const float target_weight = sum / 2.0f;
+  for (int i = 0; i < num_items; ++i) {
+    if (weighted_deltas[i].weight > 0.0f) {
+      current_weight += weighted_deltas[i].weight;
+      if (current_weight >= target_weight) {
+        return weighted_deltas[i].delta;
+      }
+    }
+  }
+  LOGW("Median not found! %d points, sum of %.2f", num_items, sum);
+  return 0.0f;
+}
+
+Point2f FramePair::GetWeightedMedian(
+    const float* const weights, const Point2f* const deltas) const {
+  Point2f median_delta;
+
+  // TODO(andrewharp): only sort deltas that could possibly have an effect.
+  static WeightedDelta weighted_deltas[kMaxKeypoints];
+
+  // Compute median X value.
+  {
+    float total_weight = 0.0f;
+
+    // Compute weighted mean and deltas.
+    for (int i = 0; i < kMaxKeypoints; ++i) {
+      weighted_deltas[i].delta = deltas[i].x;
+      const float weight = weights[i];
+      weighted_deltas[i].weight = weight;
+      if (weight > 0.0f) {
+        total_weight += weight;
+      }
+    }
+    qsort(weighted_deltas, kMaxKeypoints, sizeof(WeightedDelta),
+          WeightedDeltaCompare);
+    median_delta.x = GetMedian(kMaxKeypoints, weighted_deltas, total_weight);
+  }
+
+  // Compute median Y value.
+  {
+    float total_weight = 0.0f;
+
+    // Compute weighted mean and deltas.
+    for (int i = 0; i < kMaxKeypoints; ++i) {
+      const float weight = weights[i];
+      weighted_deltas[i].weight = weight;
+      weighted_deltas[i].delta = deltas[i].y;
+      if (weight > 0.0f) {
+        total_weight += weight;
+      }
+    }
+    qsort(weighted_deltas, kMaxKeypoints, sizeof(WeightedDelta),
+          WeightedDeltaCompare);
+    median_delta.y = GetMedian(kMaxKeypoints, weighted_deltas, total_weight);
+  }
+
+  return median_delta;
+}
+
+float FramePair::GetWeightedMedianScale(
+    const float* const weights, const Point2f* const deltas) const {
+  float median_delta;
+
+  // TODO(andrewharp): only sort deltas that could possibly have an effect.
+  static WeightedDelta weighted_deltas[kMaxKeypoints * 2];
+
+  // Compute median scale value across x and y.
+  {
+    float total_weight = 0.0f;
+
+    // Add X values.
+    for (int i = 0; i < kMaxKeypoints; ++i) {
+      weighted_deltas[i].delta = deltas[i].x;
+      const float weight = weights[i];
+      weighted_deltas[i].weight = weight;
+      if (weight > 0.0f) {
+        total_weight += weight;
+      }
+    }
+
+    // Add Y values.
+    for (int i = 0; i < kMaxKeypoints; ++i) {
+      weighted_deltas[i + kMaxKeypoints].delta = deltas[i].y;
+      const float weight = weights[i];
+      weighted_deltas[i + kMaxKeypoints].weight = weight;
+      if (weight > 0.0f) {
+        total_weight += weight;
+      }
+    }
+
+    qsort(weighted_deltas, kMaxKeypoints * 2, sizeof(WeightedDelta),
+          WeightedDeltaCompare);
+
+    median_delta = GetMedian(kMaxKeypoints * 2, weighted_deltas, total_weight);
+  }
+
+  return median_delta;
+}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/frame_pair.h

@@ -0,0 +1,104 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_FRAME_PAIR_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_FRAME_PAIR_H_
+
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint.h"
+
+namespace tf_tracking {
+
+// A class that records keypoint correspondences from pairs of
+// consecutive frames.
+class FramePair {
+ public:
+  FramePair()
+      : start_time_(0),
+        end_time_(0),
+        number_of_keypoints_(0) {}
+
+  // Cleans up the FramePair so that they can be reused.
+  void Init(const int64_t start_time, const int64_t end_time);
+
+  void AdjustBox(const BoundingBox box,
+                 float* const translation_x,
+                 float* const translation_y,
+                 float* const scale_x,
+                 float* const scale_y) const;
+
+ private:
+  // Returns the weighted median of the given deltas, computed independently on
+  // x and y. Returns 0,0 in case of failure. The assumption is that a
+  // translation of 0.0 in the degenerate case is the best that can be done, and
+  // should not be considered an error.
+  //
+  // In the case of scale,  a slight exception is made just to be safe and
+  // there is a check for 0.0 explicitly, but that shouldn't ever be possible to
+  // happen naturally because of the non-zero + parity checks in FillScales.
+  Point2f GetWeightedMedian(const float* const weights,
+                            const Point2f* const deltas) const;
+
+  float GetWeightedMedianScale(const float* const weights,
+                               const Point2f* const deltas) const;
+
+  // Weights points based on the query_point and cutoff_dist.
+  int FillWeights(const BoundingBox& box,
+                  float* const weights) const;
+
+  // Fills in the array of deltas with the translations of the points
+  // between frames.
+  void FillTranslations(Point2f* const translations) const;
+
+  // Fills in the array of deltas with the relative scale factor of points
+  // relative to a given center. Has the ability to override the weight to 0 if
+  // a degenerate scale is detected.
+  // Translation is the amount the center of the box has moved from one frame to
+  // the next.
+  int FillScales(const Point2f& old_center,
+                 const Point2f& translation,
+                 float* const weights,
+                 Point2f* const scales) const;
+
+  // TODO(andrewharp): Make these private.
+ public:
+  // The time at frame1.
+  int64_t start_time_;
+
+  // The time at frame2.
+  int64_t end_time_;
+
+  // This array will contain the keypoints found in frame 1.
+  Keypoint frame1_keypoints_[kMaxKeypoints];
+
+  // Contain the locations of the keypoints from frame 1 in frame 2.
+  Keypoint frame2_keypoints_[kMaxKeypoints];
+
+  // The number of keypoints in frame 1.
+  int number_of_keypoints_;
+
+  // Keeps track of which keypoint correspondences were actually found from one
+  // frame to another.
+  // The i-th element of this array will be non-zero if and only if the i-th
+  // keypoint of frame 1 was found in frame 2.
+  bool optical_flow_found_keypoint_[kMaxKeypoints];
+
+ private:
+  FramePair(const FramePair&) = delete;
+  void operator=(const FramePair&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_FRAME_PAIR_H_

tensorflow/tools/android/test/jni/object_tracking/geom.h

@@ -0,0 +1,319 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_GEOM_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_GEOM_H_
+
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+struct Size {
+  Size(const int width, const int height) : width(width), height(height) {}
+
+  int width;
+  int height;
+};
+
+
+class Point2f {
+ public:
+  Point2f() : x(0.0f), y(0.0f) {}
+  Point2f(const float x, const float y) : x(x), y(y) {}
+
+  inline Point2f operator- (const Point2f& that) const {
+    return Point2f(this->x - that.x, this->y - that.y);
+  }
+
+  inline Point2f operator+ (const Point2f& that) const {
+    return Point2f(this->x + that.x, this->y + that.y);
+  }
+
+  inline Point2f& operator+= (const Point2f& that) {
+    this->x += that.x;
+    this->y += that.y;
+    return *this;
+  }
+
+  inline Point2f& operator-= (const Point2f& that) {
+    this->x -= that.x;
+    this->y -= that.y;
+    return *this;
+  }
+
+  inline Point2f operator- (const Point2f& that) {
+    return Point2f(this->x - that.x, this->y - that.y);
+  }
+
+  inline float LengthSquared() {
+    return Square(this->x) + Square(this->y);
+  }
+
+  inline float Length() {
+    return sqrtf(LengthSquared());
+  }
+
+  inline float DistanceSquared(const Point2f& that) {
+    return Square(this->x - that.x) + Square(this->y - that.y);
+  }
+
+  inline float Distance(const Point2f& that) {
+    return sqrtf(DistanceSquared(that));
+  }
+
+  float x;
+  float y;
+};
+
+inline std::ostream& operator<<(std::ostream& stream, const Point2f& point) {
+  stream << point.x << "," << point.y;
+  return stream;
+}
+
+class BoundingBox {
+ public:
+  BoundingBox()
+      : left_(0),
+        top_(0),
+        right_(0),
+        bottom_(0) {}
+
+  BoundingBox(const BoundingBox& bounding_box)
+      : left_(bounding_box.left_),
+        top_(bounding_box.top_),
+        right_(bounding_box.right_),
+        bottom_(bounding_box.bottom_) {
+    SCHECK(left_ < right_, "Bounds out of whack! %.2f vs %.2f!", left_, right_);
+    SCHECK(top_ < bottom_, "Bounds out of whack! %.2f vs %.2f!", top_, bottom_);
+  }
+
+  BoundingBox(const float left,
+              const float top,
+              const float right,
+              const float bottom)
+      : left_(left),
+        top_(top),
+        right_(right),
+        bottom_(bottom) {
+    SCHECK(left_ < right_, "Bounds out of whack! %.2f vs %.2f!", left_, right_);
+    SCHECK(top_ < bottom_, "Bounds out of whack! %.2f vs %.2f!", top_, bottom_);
+  }
+
+  BoundingBox(const Point2f& point1, const Point2f& point2)
+      : left_(MIN(point1.x, point2.x)),
+        top_(MIN(point1.y, point2.y)),
+        right_(MAX(point1.x, point2.x)),
+        bottom_(MAX(point1.y, point2.y)) {}
+
+  inline void CopyToArray(float* const bounds_array) const {
+    bounds_array[0] = left_;
+    bounds_array[1] = top_;
+    bounds_array[2] = right_;
+    bounds_array[3] = bottom_;
+  }
+
+  inline float GetWidth() const {
+    return right_ - left_;
+  }
+
+  inline float GetHeight() const {
+    return bottom_ - top_;
+  }
+
+  inline float GetArea() const {
+    const float width = GetWidth();
+    const float height = GetHeight();
+    if (width <= 0 || height <= 0) {
+      return 0.0f;
+    }
+
+    return width * height;
+  }
+
+  inline Point2f GetCenter() const {
+    return Point2f((left_ + right_) / 2.0f,
+                   (top_ + bottom_) / 2.0f);
+  }
+
+  inline bool ValidBox() const {
+    return GetArea() > 0.0f;
+  }
+
+  // Returns a bounding box created from the overlapping area of these two.
+  inline BoundingBox Intersect(const BoundingBox& that) const {
+    const float new_left = MAX(this->left_, that.left_);
+    const float new_right = MIN(this->right_, that.right_);
+
+    if (new_left >= new_right) {
+      return BoundingBox();
+    }
+
+    const float new_top = MAX(this->top_, that.top_);
+    const float new_bottom = MIN(this->bottom_, that.bottom_);
+
+    if (new_top >= new_bottom) {
+      return BoundingBox();
+    }
+
+    return BoundingBox(new_left, new_top,  new_right, new_bottom);
+  }
+
+  // Returns a bounding box that can contain both boxes.
+  inline BoundingBox Union(const BoundingBox& that) const {
+    return BoundingBox(MIN(this->left_, that.left_),
+                       MIN(this->top_, that.top_),
+                       MAX(this->right_, that.right_),
+                       MAX(this->bottom_, that.bottom_));
+  }
+
+  inline float PascalScore(const BoundingBox& that) const {
+    SCHECK(GetArea() > 0.0f, "Empty bounding box!");
+    SCHECK(that.GetArea() > 0.0f, "Empty bounding box!");
+
+    const float intersect_area = this->Intersect(that).GetArea();
+
+    if (intersect_area <= 0) {
+      return 0;
+    }
+
+    const float score =
+        intersect_area / (GetArea() + that.GetArea() - intersect_area);
+    SCHECK(InRange(score, 0.0f, 1.0f), "Invalid score! %.2f", score);
+    return score;
+  }
+
+  inline bool Intersects(const BoundingBox& that) const {
+    return InRange(that.left_, left_, right_)
+        || InRange(that.right_, left_, right_)
+        || InRange(that.top_, top_, bottom_)
+        || InRange(that.bottom_, top_, bottom_);
+  }
+
+  // Returns whether another bounding box is completely inside of this bounding
+  // box. Sharing edges is ok.
+  inline bool Contains(const BoundingBox& that) const {
+    return that.left_ >= left_ &&
+        that.right_ <= right_ &&
+        that.top_ >= top_ &&
+        that.bottom_ <= bottom_;
+  }
+
+  inline bool Contains(const Point2f& point) const {
+    return InRange(point.x, left_, right_) && InRange(point.y, top_, bottom_);
+  }
+
+  inline void Shift(const Point2f shift_amount) {
+    left_ += shift_amount.x;
+    top_ += shift_amount.y;
+    right_ += shift_amount.x;
+    bottom_ += shift_amount.y;
+  }
+
+  inline void ScaleOrigin(const float scale_x, const float scale_y) {
+    left_ *= scale_x;
+    right_ *= scale_x;
+    top_ *= scale_y;
+    bottom_ *= scale_y;
+  }
+
+  inline void Scale(const float scale_x, const float scale_y) {
+    const Point2f center = GetCenter();
+    const float half_width = GetWidth() / 2.0f;
+    const float half_height = GetHeight() / 2.0f;
+
+    left_ = center.x - half_width * scale_x;
+    right_ = center.x + half_width * scale_x;
+
+    top_ = center.y - half_height * scale_y;
+    bottom_ = center.y + half_height * scale_y;
+  }
+
+  float left_;
+  float top_;
+  float right_;
+  float bottom_;
+};
+inline std::ostream& operator<<(std::ostream& stream, const BoundingBox& box) {
+  stream << "[" << box.left_ << " - " << box.right_
+         << ", " << box.top_ << " - " << box.bottom_
+         << ",  w:" << box.GetWidth() << " h:" << box.GetHeight() << "]";
+  return stream;
+}
+
+
+class BoundingSquare {
+ public:
+  BoundingSquare(const float x, const float y, const float size)
+      : x_(x), y_(y), size_(size) {}
+
+  explicit BoundingSquare(const BoundingBox& box)
+      : x_(box.left_), y_(box.top_), size_(box.GetWidth()) {
+#ifdef SANITY_CHECKS
+    if (std::abs(box.GetWidth() - box.GetHeight()) > 0.1f) {
+      LOG(WARNING) << "This is not a square: " << box << std::endl;
+    }
+#endif
+  }
+
+  inline BoundingBox ToBoundingBox() const {
+    return BoundingBox(x_, y_, x_ + size_, y_ + size_);
+  }
+
+  inline bool ValidBox() {
+    return size_ > 0.0f;
+  }
+
+  inline void Shift(const Point2f shift_amount) {
+    x_ += shift_amount.x;
+    y_ += shift_amount.y;
+  }
+
+  inline void Scale(const float scale) {
+    const float new_size = size_ * scale;
+    const float position_diff = (new_size - size_) / 2.0f;
+    x_ -= position_diff;
+    y_ -= position_diff;
+    size_ = new_size;
+  }
+
+  float x_;
+  float y_;
+  float size_;
+};
+inline std::ostream& operator<<(std::ostream& stream,
+                                const BoundingSquare& square) {
+  stream << "[" << square.x_ << "," << square.y_ << " " << square.size_ << "]";
+  return stream;
+}
+
+
+inline BoundingSquare GetCenteredSquare(const BoundingBox& original_box,
+                                        const float size) {
+  const float width_diff = (original_box.GetWidth() - size) / 2.0f;
+  const float height_diff = (original_box.GetHeight() - size) / 2.0f;
+  return BoundingSquare(original_box.left_ + width_diff,
+                        original_box.top_ + height_diff,
+                        size);
+}
+
+inline BoundingSquare GetCenteredSquare(const BoundingBox& original_box) {
+  return GetCenteredSquare(
+      original_box, MIN(original_box.GetWidth(), original_box.GetHeight()));
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_GEOM_H_

tensorflow/tools/android/test/jni/object_tracking/gl_utils.h

@@ -0,0 +1,55 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_GL_UTILS_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_GL_UTILS_H_
+
+#include <GLES/gl.h>
+#include <GLES/glext.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+
+namespace tf_tracking {
+
+// Draws a box at the given position.
+inline static void DrawBox(const BoundingBox& bounding_box) {
+  const GLfloat line[] = {
+      bounding_box.left_, bounding_box.bottom_,
+      bounding_box.left_, bounding_box.top_,
+      bounding_box.left_, bounding_box.top_,
+      bounding_box.right_, bounding_box.top_,
+      bounding_box.right_, bounding_box.top_,
+      bounding_box.right_, bounding_box.bottom_,
+      bounding_box.right_, bounding_box.bottom_,
+      bounding_box.left_, bounding_box.bottom_
+  };
+
+  glVertexPointer(2, GL_FLOAT, 0, line);
+  glEnableClientState(GL_VERTEX_ARRAY);
+
+  glDrawArrays(GL_LINES, 0, 8);
+}
+
+
+// Changes the coordinate system such that drawing to an arbitrary square in
+// the world can thereafter be drawn to using coordinates 0 - 1.
+inline static void MapWorldSquareToUnitSquare(const BoundingSquare& square) {
+  glScalef(square.size_, square.size_, 1.0f);
+  glTranslatef(square.x_ / square.size_, square.y_ / square.size_, 0.0f);
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_GL_UTILS_H_

tensorflow/tools/android/test/jni/object_tracking/image-inl.h

@@ -0,0 +1,644 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_INL_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_INL_H_
+
+#include <stdint.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+template <typename T>
+Image<T>::Image(const int width, const int height)
+    : width_less_one_(width - 1),
+      height_less_one_(height - 1),
+      data_size_(width * height),
+      own_data_(true),
+      width_(width),
+      height_(height),
+      stride_(width) {
+  Allocate();
+}
+
+template <typename T>
+Image<T>::Image(const Size& size)
+    : width_less_one_(size.width - 1),
+      height_less_one_(size.height - 1),
+      data_size_(size.width * size.height),
+      own_data_(true),
+      width_(size.width),
+      height_(size.height),
+      stride_(size.width) {
+  Allocate();
+}
+
+// Constructor that creates an image from preallocated data.
+// Note: The image takes ownership of the data lifecycle, unless own_data is
+// set to false.
+template <typename T>
+Image<T>::Image(const int width, const int height, T* const image_data,
+      const bool own_data) :
+    width_less_one_(width - 1),
+    height_less_one_(height - 1),
+    data_size_(width * height),
+    own_data_(own_data),
+    width_(width),
+    height_(height),
+    stride_(width) {
+  image_data_ = image_data;
+  SCHECK(image_data_ != NULL, "Can't create image with NULL data!");
+}
+
+template <typename T>
+Image<T>::~Image() {
+  if (own_data_) {
+    delete[] image_data_;
+  }
+  image_data_ = NULL;
+}
+
+template<typename T>
+template<class DstType>
+bool Image<T>::ExtractPatchAtSubpixelFixed1616(const int fp_x,
+                                               const int fp_y,
+                                               const int patchwidth,
+                                               const int patchheight,
+                                               DstType* to_data) const {
+  // Calculate weights.
+  const int trunc_x = fp_x >> 16;
+  const int trunc_y = fp_y >> 16;
+
+  if (trunc_x < 0 || trunc_y < 0 ||
+      (trunc_x + patchwidth) >= width_less_one_ ||
+      (trunc_y + patchheight) >= height_less_one_) {
+    return false;
+  }
+
+  // Now walk over destination patch and fill from interpolated source image.
+  for (int y = 0; y < patchheight; ++y, to_data += patchwidth) {
+    for (int x = 0; x < patchwidth; ++x) {
+      to_data[x] =
+          static_cast<DstType>(GetPixelInterpFixed1616(fp_x + (x << 16),
+                                                       fp_y + (y << 16)));
+    }
+  }
+
+  return true;
+}
+
+template <typename T>
+Image<T>* Image<T>::Crop(
+    const int left, const int top, const int right, const int bottom) const {
+  SCHECK(left >= 0 && left < width_, "out of bounds at %d!", left);
+  SCHECK(right >= 0 && right < width_, "out of bounds at %d!", right);
+  SCHECK(top >= 0 && top < height_, "out of bounds at %d!", top);
+  SCHECK(bottom >= 0 && bottom < height_, "out of bounds at %d!", bottom);
+
+  SCHECK(left <= right, "mismatch!");
+  SCHECK(top <= bottom, "mismatch!");
+
+  const int new_width = right - left + 1;
+  const int new_height = bottom - top + 1;
+
+  Image<T>* const cropped_image = new Image(new_width, new_height);
+
+  for (int y = 0; y < new_height; ++y) {
+    memcpy((*cropped_image)[y], ((*this)[y + top] + left),
+           new_width * sizeof(T));
+  }
+
+  return cropped_image;
+}
+
+template <typename T>
+inline float Image<T>::GetPixelInterp(const float x, const float y) const {
+  // Do int conversion one time.
+  const int floored_x = static_cast<int>(x);
+  const int floored_y = static_cast<int>(y);
+
+  // Note: it might be the case that the *_[min|max] values are clipped, and
+  // these (the a b c d vals) aren't (for speed purposes), but that doesn't
+  // matter. We'll just be blending the pixel with itself in that case anyway.
+  const float b = x - floored_x;
+  const float a = 1.0f - b;
+
+  const float d = y - floored_y;
+  const float c = 1.0f - d;
+
+  SCHECK(ValidInterpPixel(x, y),
+        "x or y out of bounds! %.2f [0 - %d), %.2f [0 - %d)",
+        x, width_less_one_, y, height_less_one_);
+
+  const T* const pix_ptr = (*this)[floored_y] + floored_x;
+
+  // Get the pixel values surrounding this point.
+  const T& p1 = pix_ptr[0];
+  const T& p2 = pix_ptr[1];
+  const T& p3 = pix_ptr[width_];
+  const T& p4 = pix_ptr[width_ + 1];
+
+  // Simple bilinear interpolation between four reference pixels.
+  // If x is the value requested:
+  //     a  b
+  //   -------
+  // c |p1 p2|
+  //   |  x  |
+  // d |p3 p4|
+  //   -------
+  return  c * ((a * p1) + (b * p2)) +
+          d * ((a * p3) + (b * p4));
+}
+
+
+template <typename T>
+inline T Image<T>::GetPixelInterpFixed1616(
+    const int fp_x_whole, const int fp_y_whole) const {
+  static const int kFixedPointOne = 0x00010000;
+  static const int kFixedPointHalf = 0x00008000;
+  static const int kFixedPointTruncateMask = 0xFFFF0000;
+
+  int trunc_x = fp_x_whole & kFixedPointTruncateMask;
+  int trunc_y = fp_y_whole & kFixedPointTruncateMask;
+  const int fp_x = fp_x_whole - trunc_x;
+  const int fp_y = fp_y_whole - trunc_y;
+
+  // Scale the truncated values back to regular ints.
+  trunc_x >>= 16;
+  trunc_y >>= 16;
+
+  const int one_minus_fp_x = kFixedPointOne - fp_x;
+  const int one_minus_fp_y = kFixedPointOne - fp_y;
+
+  const T* trunc_start = (*this)[trunc_y] + trunc_x;
+
+  const T a = trunc_start[0];
+  const T b = trunc_start[1];
+  const T c = trunc_start[stride_];
+  const T d = trunc_start[stride_ + 1];
+
+  return (
+      (one_minus_fp_y * static_cast<int64_t>(one_minus_fp_x * a + fp_x * b) +
+       fp_y * static_cast<int64_t>(one_minus_fp_x * c + fp_x * d) +
+       kFixedPointHalf) >>
+      32);
+}
+
+template <typename T>
+inline bool Image<T>::ValidPixel(const int x, const int y) const {
+  return InRange(x, ZERO, width_less_one_) &&
+         InRange(y, ZERO, height_less_one_);
+}
+
+template <typename T>
+inline BoundingBox Image<T>::GetContainingBox() const {
+  return BoundingBox(
+      0, 0, width_less_one_ - EPSILON, height_less_one_ - EPSILON);
+}
+
+template <typename T>
+inline bool Image<T>::Contains(const BoundingBox& bounding_box) const {
+  // TODO(andrewharp): Come up with a more elegant way of ensuring that bounds
+  // are ok.
+  return GetContainingBox().Contains(bounding_box);
+}
+
+template <typename T>
+inline bool Image<T>::ValidInterpPixel(const float x, const float y) const {
+  // Exclusive of max because we can be more efficient if we don't handle
+  // interpolating on or past the last pixel.
+  return (x >= ZERO) && (x < width_less_one_) &&
+         (y >= ZERO) && (y < height_less_one_);
+}
+
+template <typename T>
+void Image<T>::DownsampleAveraged(const T* const original, const int stride,
+                                  const int factor) {
+#ifdef __ARM_NEON
+  if (factor == 4 || factor == 2) {
+    DownsampleAveragedNeon(original, stride, factor);
+    return;
+  }
+#endif
+
+  // TODO(andrewharp): delete or enable this for non-uint8_t downsamples.
+  const int pixels_per_block = factor * factor;
+
+  // For every pixel in resulting image.
+  for (int y = 0; y < height_; ++y) {
+    const int orig_y = y * factor;
+    const int y_bound = orig_y + factor;
+
+    // Sum up the original pixels.
+    for (int x = 0; x < width_; ++x) {
+      const int orig_x = x * factor;
+      const int x_bound = orig_x + factor;
+
+      // Making this int32_t because type U or T might overflow.
+      int32_t pixel_sum = 0;
+
+      // Grab all the pixels that make up this pixel.
+      for (int curr_y = orig_y; curr_y < y_bound; ++curr_y) {
+        const T* p = original + curr_y * stride + orig_x;
+
+        for (int curr_x = orig_x; curr_x < x_bound; ++curr_x) {
+          pixel_sum += *p++;
+        }
+      }
+
+      (*this)[y][x] = pixel_sum / pixels_per_block;
+    }
+  }
+}
+
+template <typename T>
+void Image<T>::DownsampleInterpolateNearest(const Image<T>& original) {
+  // Calculating the scaling factors based on target image size.
+  const float factor_x = static_cast<float>(original.GetWidth()) /
+      static_cast<float>(width_);
+  const float factor_y = static_cast<float>(original.GetHeight()) /
+      static_cast<float>(height_);
+
+  // Calculating initial offset in x-axis.
+  const float offset_x = 0.5f * (original.GetWidth() - width_) / width_;
+
+  // Calculating initial offset in y-axis.
+  const float offset_y = 0.5f * (original.GetHeight() - height_) / height_;
+
+  float orig_y = offset_y;
+
+  // For every pixel in resulting image.
+  for (int y = 0; y < height_; ++y) {
+    float orig_x = offset_x;
+
+    // Finding nearest pixel on y-axis.
+    const int nearest_y = static_cast<int>(orig_y + 0.5f);
+    const T* row_data = original[nearest_y];
+
+    T* pixel_ptr = (*this)[y];
+
+    for (int x = 0; x < width_; ++x) {
+      // Finding nearest pixel on x-axis.
+      const int nearest_x = static_cast<int>(orig_x + 0.5f);
+
+      *pixel_ptr++ = row_data[nearest_x];
+
+      orig_x += factor_x;
+    }
+
+    orig_y += factor_y;
+  }
+}
+
+template <typename T>
+void Image<T>::DownsampleInterpolateLinear(const Image<T>& original) {
+  // TODO(andrewharp): Turn this into a general compare sizes/bulk
+  // copy method.
+  if (original.GetWidth() == GetWidth() &&
+      original.GetHeight() == GetHeight() &&
+      original.stride() == stride()) {
+    memcpy(image_data_, original.data(), data_size_ * sizeof(T));
+    return;
+  }
+
+  // Calculating the scaling factors based on target image size.
+  const float factor_x = static_cast<float>(original.GetWidth()) /
+      static_cast<float>(width_);
+  const float factor_y = static_cast<float>(original.GetHeight()) /
+      static_cast<float>(height_);
+
+  // Calculating initial offset in x-axis.
+  const float offset_x = 0;
+  const int offset_x_fp = RealToFixed1616(offset_x);
+
+  // Calculating initial offset in y-axis.
+  const float offset_y = 0;
+  const int offset_y_fp = RealToFixed1616(offset_y);
+
+  // Get the fixed point scaling factor value.
+  // Shift by 8 so we can fit everything into a 4 byte int later for speed
+  // reasons. This means the precision is limited to 1 / 256th of a pixel,
+  // but this should be good enough.
+  const int factor_x_fp = RealToFixed1616(factor_x) >> 8;
+  const int factor_y_fp = RealToFixed1616(factor_y) >> 8;
+
+  int src_y_fp = offset_y_fp >> 8;
+
+  static const int kFixedPointOne8 = 0x00000100;
+  static const int kFixedPointHalf8 = 0x00000080;
+  static const int kFixedPointTruncateMask8 = 0xFFFFFF00;
+
+  // For every pixel in resulting image.
+  for (int y = 0; y < height_; ++y) {
+    int src_x_fp = offset_x_fp >> 8;
+
+    int trunc_y = src_y_fp & kFixedPointTruncateMask8;
+    const int fp_y = src_y_fp - trunc_y;
+
+    // Scale the truncated values back to regular ints.
+    trunc_y >>= 8;
+
+    const int one_minus_fp_y = kFixedPointOne8 - fp_y;
+
+    T* pixel_ptr = (*this)[y];
+
+    // Make sure not to read from an invalid row.
+    const int trunc_y_b = MIN(original.height_less_one_, trunc_y + 1);
+    const T* other_top_ptr = original[trunc_y];
+    const T* other_bot_ptr = original[trunc_y_b];
+
+    int last_trunc_x = -1;
+    int trunc_x = -1;
+
+    T a = 0;
+    T b = 0;
+    T c = 0;
+    T d = 0;
+
+    for (int x = 0; x < width_; ++x) {
+      trunc_x = src_x_fp & kFixedPointTruncateMask8;
+
+      const int fp_x = (src_x_fp - trunc_x) >> 8;
+
+      // Scale the truncated values back to regular ints.
+      trunc_x >>= 8;
+
+      // It's possible we're reading from the same pixels
+      if (trunc_x != last_trunc_x) {
+        // Make sure not to read from an invalid column.
+        const int trunc_x_b = MIN(original.width_less_one_, trunc_x + 1);
+        a = other_top_ptr[trunc_x];
+        b = other_top_ptr[trunc_x_b];
+        c = other_bot_ptr[trunc_x];
+        d = other_bot_ptr[trunc_x_b];
+        last_trunc_x = trunc_x;
+      }
+
+      const int one_minus_fp_x = kFixedPointOne8 - fp_x;
+
+      const int32_t value =
+          ((one_minus_fp_y * one_minus_fp_x * a + fp_x * b) +
+           (fp_y * one_minus_fp_x * c + fp_x * d) + kFixedPointHalf8) >>
+          16;
+
+      *pixel_ptr++ = value;
+
+      src_x_fp += factor_x_fp;
+    }
+    src_y_fp += factor_y_fp;
+  }
+}
+
+template <typename T>
+void Image<T>::DownsampleSmoothed3x3(const Image<T>& original) {
+  for (int y = 0; y < height_; ++y) {
+    const int orig_y = Clip(2 * y, ZERO, original.height_less_one_);
+    const int min_y = Clip(orig_y - 1, ZERO, original.height_less_one_);
+    const int max_y = Clip(orig_y + 1, ZERO, original.height_less_one_);
+
+    for (int x = 0; x < width_; ++x) {
+      const int orig_x = Clip(2 * x, ZERO, original.width_less_one_);
+      const int min_x = Clip(orig_x - 1, ZERO, original.width_less_one_);
+      const int max_x = Clip(orig_x + 1, ZERO, original.width_less_one_);
+
+      // Center.
+      int32_t pixel_sum = original[orig_y][orig_x] * 4;
+
+      // Sides.
+      pixel_sum += (original[orig_y][max_x] +
+                    original[orig_y][min_x] +
+                    original[max_y][orig_x] +
+                    original[min_y][orig_x]) * 2;
+
+      // Diagonals.
+      pixel_sum += (original[min_y][max_x] +
+                    original[min_y][min_x] +
+                    original[max_y][max_x] +
+                    original[max_y][min_x]);
+
+      (*this)[y][x] = pixel_sum >> 4;  // 16
+    }
+  }
+}
+
+template <typename T>
+void Image<T>::DownsampleSmoothed5x5(const Image<T>& original) {
+  const int max_x = original.width_less_one_;
+  const int max_y = original.height_less_one_;
+
+  // The JY Bouget paper on Lucas-Kanade recommends a
+  // [1/16 1/4 3/8 1/4 1/16]^2 filter.
+  // This works out to a [1 4 6 4 1]^2 / 256 array, precomputed below.
+  static const int window_radius = 2;
+  static const int window_size = window_radius*2 + 1;
+  static const int window_weights[] = {1,  4,  6,  4, 1,   // 16 +
+                                       4, 16, 24, 16, 4,   // 64 +
+                                       6, 24, 36, 24, 6,   // 96 +
+                                       4, 16, 24, 16, 4,   // 64 +
+                                       1,  4,  6,  4, 1};  // 16 = 256
+
+  // We'll multiply and sum with the whole numbers first, then divide by
+  // the total weight to normalize at the last moment.
+  for (int y = 0; y < height_; ++y) {
+    for (int x = 0; x < width_; ++x) {
+      int32_t pixel_sum = 0;
+
+      const int* w = window_weights;
+      const int start_x = Clip((x << 1) - window_radius, ZERO, max_x);
+
+      // Clip the boundaries to the size of the image.
+      for (int window_y = 0; window_y < window_size; ++window_y) {
+        const int start_y =
+            Clip((y << 1) - window_radius + window_y, ZERO, max_y);
+
+        const T* p = original[start_y] + start_x;
+
+        for (int window_x = 0; window_x < window_size; ++window_x) {
+          pixel_sum +=  *p++ * *w++;
+        }
+      }
+
+      // Conversion to type T will happen here after shifting right 8 bits to
+      // divide by 256.
+      (*this)[y][x] = pixel_sum >> 8;
+    }
+  }
+}
+
+template <typename T>
+template <typename U>
+inline T Image<T>::ScharrPixelX(const Image<U>& original,
+                      const int center_x, const int center_y) const {
+  const int min_x = Clip(center_x - 1, ZERO, original.width_less_one_);
+  const int max_x = Clip(center_x + 1, ZERO, original.width_less_one_);
+  const int min_y = Clip(center_y - 1, ZERO, original.height_less_one_);
+  const int max_y = Clip(center_y + 1, ZERO, original.height_less_one_);
+
+  // Convolution loop unrolled for performance...
+  return (3 * (original[min_y][max_x]
+               + original[max_y][max_x]
+               - original[min_y][min_x]
+               - original[max_y][min_x])
+          + 10 * (original[center_y][max_x]
+                  - original[center_y][min_x])) / 32;
+}
+
+template <typename T>
+template <typename U>
+inline T Image<T>::ScharrPixelY(const Image<U>& original,
+                      const int center_x, const int center_y) const {
+  const int min_x = Clip(center_x - 1, 0, original.width_less_one_);
+  const int max_x = Clip(center_x + 1, 0, original.width_less_one_);
+  const int min_y = Clip(center_y - 1, 0, original.height_less_one_);
+  const int max_y = Clip(center_y + 1, 0, original.height_less_one_);
+
+  // Convolution loop unrolled for performance...
+  return (3 * (original[max_y][min_x]
+               + original[max_y][max_x]
+               - original[min_y][min_x]
+               - original[min_y][max_x])
+          + 10 * (original[max_y][center_x]
+                  - original[min_y][center_x])) / 32;
+}
+
+template <typename T>
+template <typename U>
+inline void Image<T>::ScharrX(const Image<U>& original) {
+  for (int y = 0; y < height_; ++y) {
+    for (int x = 0; x < width_; ++x) {
+      SetPixel(x, y, ScharrPixelX(original, x, y));
+    }
+  }
+}
+
+template <typename T>
+template <typename U>
+inline void Image<T>::ScharrY(const Image<U>& original) {
+  for (int y = 0; y < height_; ++y) {
+    for (int x = 0; x < width_; ++x) {
+      SetPixel(x, y, ScharrPixelY(original, x, y));
+    }
+  }
+}
+
+template <typename T>
+template <typename U>
+void Image<T>::DerivativeX(const Image<U>& original) {
+  for (int y = 0; y < height_; ++y) {
+    const U* const source_row = original[y];
+    T* const dest_row = (*this)[y];
+
+    // Compute first pixel. Approximated with forward difference.
+    dest_row[0] = source_row[1] - source_row[0];
+
+    // All the pixels in between. Central difference method.
+    const U* source_prev_pixel = source_row;
+    T* dest_pixel = dest_row + 1;
+    const U* source_next_pixel = source_row + 2;
+    for (int x = 1; x < width_less_one_; ++x) {
+      *dest_pixel++ = HalfDiff(*source_prev_pixel++, *source_next_pixel++);
+    }
+
+    // Last pixel. Approximated with backward difference.
+    dest_row[width_less_one_] =
+        source_row[width_less_one_] - source_row[width_less_one_ - 1];
+  }
+}
+
+template <typename T>
+template <typename U>
+void Image<T>::DerivativeY(const Image<U>& original) {
+  const int src_stride = original.stride();
+
+  // Compute 1st row. Approximated with forward difference.
+  {
+    const U* const src_row = original[0];
+    T* dest_row = (*this)[0];
+    for (int x = 0; x < width_; ++x) {
+      dest_row[x] = src_row[x + src_stride] - src_row[x];
+    }
+  }
+
+  // Compute all rows in between using central difference.
+  for (int y = 1; y < height_less_one_; ++y) {
+    T* dest_row = (*this)[y];
+
+    const U* source_prev_pixel = original[y - 1];
+    const U* source_next_pixel = original[y + 1];
+    for (int x = 0; x < width_; ++x) {
+      *dest_row++ = HalfDiff(*source_prev_pixel++, *source_next_pixel++);
+    }
+  }
+
+  // Compute last row. Approximated with backward difference.
+  {
+    const U* const src_row = original[height_less_one_];
+    T* dest_row = (*this)[height_less_one_];
+    for (int x = 0; x < width_; ++x) {
+      dest_row[x] = src_row[x] - src_row[x - src_stride];
+    }
+  }
+}
+
+template <typename T>
+template <typename U>
+inline T Image<T>::ConvolvePixel3x3(const Image<U>& original,
+                                    const int* const filter,
+                                    const int center_x, const int center_y,
+                                    const int total) const {
+  int32_t sum = 0;
+  for (int filter_y = 0; filter_y < 3; ++filter_y) {
+    const int y = Clip(center_y - 1 + filter_y, 0, original.GetHeight());
+    for (int filter_x = 0; filter_x < 3; ++filter_x) {
+      const int x = Clip(center_x - 1 + filter_x, 0, original.GetWidth());
+      sum += original[y][x] * filter[filter_y * 3 + filter_x];
+    }
+  }
+  return sum / total;
+}
+
+template <typename T>
+template <typename U>
+inline void Image<T>::Convolve3x3(const Image<U>& original,
+                                  const int32_t* const filter) {
+  int32_t sum = 0;
+  for (int i = 0; i < 9; ++i) {
+    sum += abs(filter[i]);
+  }
+  for (int y = 0; y < height_; ++y) {
+    for (int x = 0; x < width_; ++x) {
+      SetPixel(x, y, ConvolvePixel3x3(original, filter, x, y, sum));
+    }
+  }
+}
+
+template <typename T>
+inline void Image<T>::FromArray(const T* const pixels, const int stride,
+                      const int factor) {
+  if (factor == 1 && stride == width_) {
+    // If not subsampling, memcpy per line should be faster.
+    memcpy(this->image_data_, pixels, data_size_ * sizeof(T));
+    return;
+  }
+
+  DownsampleAveraged(pixels, stride, factor);
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_INL_H_

tensorflow/tools/android/test/jni/object_tracking/image.h

@@ -0,0 +1,342 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_H_
+
+#include <stdint.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+// TODO(andrewharp): Make this a cast to uint32_t if/when we go unsigned for
+// operations.
+#define ZERO 0
+
+#ifdef SANITY_CHECKS
+  #define CHECK_PIXEL(IMAGE, X, Y) {\
+    SCHECK((IMAGE)->ValidPixel((X), (Y)), \
+          "CHECK_PIXEL(%d,%d) in %dx%d image.", \
+          static_cast<int>(X), static_cast<int>(Y), \
+          (IMAGE)->GetWidth(), (IMAGE)->GetHeight());\
+  }
+
+  #define CHECK_PIXEL_INTERP(IMAGE, X, Y) {\
+    SCHECK((IMAGE)->validInterpPixel((X), (Y)), \
+          "CHECK_PIXEL_INTERP(%.2f, %.2f) in %dx%d image.", \
+          static_cast<float>(X), static_cast<float>(Y), \
+          (IMAGE)->GetWidth(), (IMAGE)->GetHeight());\
+  }
+#else
+  #define CHECK_PIXEL(image, x, y) {}
+  #define CHECK_PIXEL_INTERP(IMAGE, X, Y) {}
+#endif
+
+namespace tf_tracking {
+
+#ifdef SANITY_CHECKS
+// Class which exists solely to provide bounds checking for array-style image
+// data access.
+template <typename T>
+class RowData {
+ public:
+  RowData(T* const row_data, const int max_col)
+      : row_data_(row_data), max_col_(max_col) {}
+
+  inline T& operator[](const int col) const {
+    SCHECK(InRange(col, 0, max_col_),
+          "Column out of range: %d (%d max)", col, max_col_);
+    return row_data_[col];
+  }
+
+  inline operator T*() const {
+    return row_data_;
+  }
+
+ private:
+  T* const row_data_;
+  const int max_col_;
+};
+#endif
+
+// Naive templated sorting function.
+template <typename T>
+int Comp(const void* a, const void* b) {
+  const T val1 = *reinterpret_cast<const T*>(a);
+  const T val2 = *reinterpret_cast<const T*>(b);
+
+  if (val1 == val2) {
+    return 0;
+  } else if (val1 < val2) {
+    return -1;
+  } else {
+    return 1;
+  }
+}
+
+// TODO(andrewharp): Make explicit which operations support negative numbers or
+// struct/class types in image data (possibly create fast multi-dim array class
+// for data where pixel arithmetic does not make sense).
+
+// Image class optimized for working on numeric arrays as grayscale image data.
+// Supports other data types as a 2D array class, so long as no pixel math
+// operations are called (convolution, downsampling, etc).
+template <typename T>
+class Image {
+ public:
+  Image(const int width, const int height);
+  explicit Image(const Size& size);
+
+  // Constructor that creates an image from preallocated data.
+  // Note: The image takes ownership of the data lifecycle, unless own_data is
+  // set to false.
+  Image(const int width, const int height, T* const image_data,
+        const bool own_data = true);
+
+  ~Image();
+
+  // Extract a pixel patch from this image, starting at a subpixel location.
+  // Uses 16:16 fixed point format for representing real values and doing the
+  // bilinear interpolation.
+  //
+  // Arguments fp_x and fp_y tell the subpixel position in fixed point format,
+  // patchwidth/patchheight give the size of the patch in pixels and
+  // to_data must be a valid pointer to a *contiguous* destination data array.
+  template<class DstType>
+  bool ExtractPatchAtSubpixelFixed1616(const int fp_x,
+                                       const int fp_y,
+                                       const int patchwidth,
+                                       const int patchheight,
+                                       DstType* to_data) const;
+
+  Image<T>* Crop(
+      const int left, const int top, const int right, const int bottom) const;
+
+  inline int GetWidth() const { return width_; }
+  inline int GetHeight() const { return height_; }
+
+  // Bilinearly sample a value between pixels.  Values must be within the image.
+  inline float GetPixelInterp(const float x, const float y) const;
+
+  // Bilinearly sample a pixels at a subpixel position using fixed point
+  // arithmetic.
+  // Avoids float<->int conversions.
+  // Values must be within the image.
+  // Arguments fp_x and fp_y tell the subpixel position in
+  // 16:16 fixed point format.
+  //
+  // Important: This function only makes sense for integer-valued images, such
+  // as Image<uint8_t> or Image<int> etc.
+  inline T GetPixelInterpFixed1616(const int fp_x_whole,
+                                   const int fp_y_whole) const;
+
+  // Returns true iff the pixel is in the image's boundaries.
+  inline bool ValidPixel(const int x, const int y) const;
+
+  inline BoundingBox GetContainingBox() const;
+
+  inline bool Contains(const BoundingBox& bounding_box) const;
+
+  inline T GetMedianValue() {
+    qsort(image_data_, data_size_, sizeof(image_data_[0]), Comp<T>);
+    return image_data_[data_size_ >> 1];
+  }
+
+  // Returns true iff the pixel is in the image's boundaries for interpolation
+  // purposes.
+  // TODO(andrewharp): check in interpolation follow-up change.
+  inline bool ValidInterpPixel(const float x, const float y) const;
+
+  // Safe lookup with boundary enforcement.
+  inline T GetPixelClipped(const int x, const int y) const {
+    return (*this)[Clip(y, ZERO, height_less_one_)]
+                  [Clip(x, ZERO, width_less_one_)];
+  }
+
+#ifdef SANITY_CHECKS
+  inline RowData<T> operator[](const int row) {
+    SCHECK(InRange(row, 0, height_less_one_),
+          "Row out of range: %d (%d max)", row, height_less_one_);
+    return RowData<T>(image_data_ + row * stride_, width_less_one_);
+  }
+
+  inline const RowData<T> operator[](const int row) const {
+    SCHECK(InRange(row, 0, height_less_one_),
+          "Row out of range: %d (%d max)", row, height_less_one_);
+    return RowData<T>(image_data_ + row * stride_, width_less_one_);
+  }
+#else
+  inline T* operator[](const int row) {
+    return image_data_ + row * stride_;
+  }
+
+  inline const T* operator[](const int row) const {
+    return image_data_ + row * stride_;
+  }
+#endif
+
+  const T* data() const { return image_data_; }
+
+  inline int stride() const { return stride_; }
+
+  // Clears image to a single value.
+  inline void Clear(const T& val) {
+    memset(image_data_, val, sizeof(*image_data_) * data_size_);
+  }
+
+#ifdef __ARM_NEON
+  void Downsample2x32ColumnsNeon(const uint8_t* const original,
+                                 const int stride, const int orig_x);
+
+  void Downsample4x32ColumnsNeon(const uint8_t* const original,
+                                 const int stride, const int orig_x);
+
+  void DownsampleAveragedNeon(const uint8_t* const original, const int stride,
+                              const int factor);
+#endif
+
+  // Naive downsampler that reduces image size by factor by averaging pixels in
+  // blocks of size factor x factor.
+  void DownsampleAveraged(const T* const original, const int stride,
+                          const int factor);
+
+  // Naive downsampler that reduces image size by factor by averaging pixels in
+  // blocks of size factor x factor.
+  inline void DownsampleAveraged(const Image<T>& original, const int factor) {
+    DownsampleAveraged(original.data(), original.GetWidth(), factor);
+  }
+
+  // Native downsampler that reduces image size using nearest interpolation
+  void DownsampleInterpolateNearest(const Image<T>& original);
+
+  // Native downsampler that reduces image size using fixed-point bilinear
+  // interpolation
+  void DownsampleInterpolateLinear(const Image<T>& original);
+
+  // Relatively efficient downsampling of an image by a factor of two with a
+  // low-pass 3x3 smoothing operation thrown in.
+  void DownsampleSmoothed3x3(const Image<T>& original);
+
+  // Relatively efficient downsampling of an image by a factor of two with a
+  // low-pass 5x5 smoothing operation thrown in.
+  void DownsampleSmoothed5x5(const Image<T>& original);
+
+  // Optimized Scharr filter on a single pixel in the X direction.
+  // Scharr filters are like central-difference operators, but have more
+  // rotational symmetry in their response because they also consider the
+  // diagonal neighbors.
+  template <typename U>
+  inline T ScharrPixelX(const Image<U>& original,
+                        const int center_x, const int center_y) const;
+
+  // Optimized Scharr filter on a single pixel in the X direction.
+  // Scharr filters are like central-difference operators, but have more
+  // rotational symmetry in their response because they also consider the
+  // diagonal neighbors.
+  template <typename U>
+  inline T ScharrPixelY(const Image<U>& original,
+                        const int center_x, const int center_y) const;
+
+  // Convolve the image with a Scharr filter in the X direction.
+  // Much faster than an equivalent generic convolution.
+  template <typename U>
+  inline void ScharrX(const Image<U>& original);
+
+  // Convolve the image with a Scharr filter in the Y direction.
+  // Much faster than an equivalent generic convolution.
+  template <typename U>
+  inline void ScharrY(const Image<U>& original);
+
+  static inline T HalfDiff(int32_t first, int32_t second) {
+    return (second - first) / 2;
+  }
+
+  template <typename U>
+  void DerivativeX(const Image<U>& original);
+
+  template <typename U>
+  void DerivativeY(const Image<U>& original);
+
+  // Generic function for convolving pixel with 3x3 filter.
+  // Filter pixels should be in row major order.
+  template <typename U>
+  inline T ConvolvePixel3x3(const Image<U>& original,
+                            const int* const filter,
+                            const int center_x, const int center_y,
+                            const int total) const;
+
+  // Generic function for convolving an image with a 3x3 filter.
+  // TODO(andrewharp): Generalize this for any size filter.
+  template <typename U>
+  inline void Convolve3x3(const Image<U>& original,
+                          const int32_t* const filter);
+
+  // Load this image's data from a data array. The data at pixels is assumed to
+  // have dimensions equivalent to this image's dimensions * factor.
+  inline void FromArray(const T* const pixels, const int stride,
+                        const int factor = 1);
+
+  // Copy the image back out to an appropriately sized data array.
+  inline void ToArray(T* const pixels) const {
+    // If not subsampling, memcpy should be faster.
+    memcpy(pixels, this->image_data_, data_size_ * sizeof(T));
+  }
+
+  // Precompute these for efficiency's sake as they're used by a lot of
+  // clipping code and loop code.
+  // TODO(andrewharp): make these only accessible by other Images.
+  const int width_less_one_;
+  const int height_less_one_;
+
+  // The raw size of the allocated data.
+  const int data_size_;
+
+ private:
+  inline void Allocate() {
+    image_data_ = new T[data_size_];
+    if (image_data_ == NULL) {
+      LOGE("Couldn't allocate image data!");
+    }
+  }
+
+  T* image_data_;
+
+  bool own_data_;
+
+  const int width_;
+  const int height_;
+
+  // The image stride (offset to next row).
+  // TODO(andrewharp): Make sure that stride is honored in all code.
+  const int stride_;
+
+  Image(const Image&) = delete;
+  void operator=(const Image&) = delete;
+};
+
+template <typename t>
+inline std::ostream& operator<<(std::ostream& stream, const Image<t>& image) {
+  for (int y = 0; y < image.GetHeight(); ++y) {
+    for (int x = 0; x < image.GetWidth(); ++x) {
+      stream << image[y][x] << " ";
+    }
+    stream << std::endl;
+  }
+  return stream;
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_H_

tensorflow/tools/android/test/jni/object_tracking/image_data.h

@@ -0,0 +1,265 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_DATA_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_DATA_H_
+
+#include <stdint.h>
+
+#include <memory>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_utils.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/integral_image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+// Class that encapsulates all bulky processed data for a frame.
+class ImageData {
+ public:
+  explicit ImageData(const int width, const int height)
+      : uv_frame_width_(width << 1),
+        uv_frame_height_(height << 1),
+        timestamp_(0),
+        image_(width, height) {
+    InitPyramid(width, height);
+    ResetComputationCache();
+  }
+
+ private:
+  void ResetComputationCache() {
+    uv_data_computed_ = false;
+    integral_image_computed_ = false;
+    for (int i = 0; i < kNumPyramidLevels; ++i) {
+      spatial_x_computed_[i] = false;
+      spatial_y_computed_[i] = false;
+      pyramid_sqrt2_computed_[i * 2] = false;
+      pyramid_sqrt2_computed_[i * 2 + 1] = false;
+    }
+  }
+
+  void InitPyramid(const int width, const int height) {
+    int level_width = width;
+    int level_height = height;
+
+    for (int i = 0; i < kNumPyramidLevels; ++i) {
+      pyramid_sqrt2_[i * 2] = NULL;
+      pyramid_sqrt2_[i * 2 + 1] = NULL;
+      spatial_x_[i] = NULL;
+      spatial_y_[i] = NULL;
+
+      level_width /= 2;
+      level_height /= 2;
+    }
+
+    // Alias the first pyramid level to image_.
+    pyramid_sqrt2_[0] = &image_;
+  }
+
+ public:
+  ~ImageData() {
+    // The first pyramid level is actually an alias to image_,
+    // so make sure it doesn't get deleted here.
+    pyramid_sqrt2_[0] = NULL;
+
+    for (int i = 0; i < kNumPyramidLevels; ++i) {
+      SAFE_DELETE(pyramid_sqrt2_[i * 2]);
+      SAFE_DELETE(pyramid_sqrt2_[i * 2 + 1]);
+      SAFE_DELETE(spatial_x_[i]);
+      SAFE_DELETE(spatial_y_[i]);
+    }
+  }
+
+  void SetData(const uint8_t* const new_frame, const int stride,
+               const int64_t timestamp, const int downsample_factor) {
+    SetData(new_frame, NULL, stride, timestamp, downsample_factor);
+  }
+
+  void SetData(const uint8_t* const new_frame, const uint8_t* const uv_frame,
+               const int stride, const int64_t timestamp,
+               const int downsample_factor) {
+    ResetComputationCache();
+
+    timestamp_ = timestamp;
+
+    TimeLog("SetData!");
+
+    pyramid_sqrt2_[0]->FromArray(new_frame, stride, downsample_factor);
+    pyramid_sqrt2_computed_[0] = true;
+    TimeLog("Downsampled image");
+
+    if (uv_frame != NULL) {
+      if (u_data_.get() == NULL) {
+        u_data_.reset(new Image<uint8_t>(uv_frame_width_, uv_frame_height_));
+        v_data_.reset(new Image<uint8_t>(uv_frame_width_, uv_frame_height_));
+      }
+
+      GetUV(uv_frame, u_data_.get(), v_data_.get());
+      uv_data_computed_ = true;
+      TimeLog("Copied UV data");
+    } else {
+      LOGV("No uv data!");
+    }
+
+#ifdef LOG_TIME
+    // If profiling is enabled, precompute here to make it easier to distinguish
+    // total costs.
+    Precompute();
+#endif
+  }
+
+  inline const uint64_t GetTimestamp() const { return timestamp_; }
+
+  inline const Image<uint8_t>* GetImage() const {
+    SCHECK(pyramid_sqrt2_computed_[0], "image not set!");
+    return pyramid_sqrt2_[0];
+  }
+
+  const Image<uint8_t>* GetPyramidSqrt2Level(const int level) const {
+    if (!pyramid_sqrt2_computed_[level]) {
+      SCHECK(level != 0, "Level equals 0!");
+      if (level == 1) {
+        const Image<uint8_t>& upper_level = *GetPyramidSqrt2Level(0);
+        if (pyramid_sqrt2_[level] == NULL) {
+          const int new_width =
+              (static_cast<int>(upper_level.GetWidth() / sqrtf(2)) + 1) / 2 * 2;
+          const int new_height =
+              (static_cast<int>(upper_level.GetHeight() / sqrtf(2)) + 1) / 2 *
+              2;
+
+          pyramid_sqrt2_[level] = new Image<uint8_t>(new_width, new_height);
+        }
+        pyramid_sqrt2_[level]->DownsampleInterpolateLinear(upper_level);
+      } else {
+        const Image<uint8_t>& upper_level = *GetPyramidSqrt2Level(level - 2);
+        if (pyramid_sqrt2_[level] == NULL) {
+          pyramid_sqrt2_[level] = new Image<uint8_t>(
+              upper_level.GetWidth() / 2, upper_level.GetHeight() / 2);
+        }
+        pyramid_sqrt2_[level]->DownsampleAveraged(
+            upper_level.data(), upper_level.stride(), 2);
+      }
+      pyramid_sqrt2_computed_[level] = true;
+    }
+    return pyramid_sqrt2_[level];
+  }
+
+  inline const Image<int32_t>* GetSpatialX(const int level) const {
+    if (!spatial_x_computed_[level]) {
+      const Image<uint8_t>& src = *GetPyramidSqrt2Level(level * 2);
+      if (spatial_x_[level] == NULL) {
+        spatial_x_[level] = new Image<int32_t>(src.GetWidth(), src.GetHeight());
+      }
+      spatial_x_[level]->DerivativeX(src);
+      spatial_x_computed_[level] = true;
+    }
+    return spatial_x_[level];
+  }
+
+  inline const Image<int32_t>* GetSpatialY(const int level) const {
+    if (!spatial_y_computed_[level]) {
+      const Image<uint8_t>& src = *GetPyramidSqrt2Level(level * 2);
+      if (spatial_y_[level] == NULL) {
+        spatial_y_[level] = new Image<int32_t>(src.GetWidth(), src.GetHeight());
+      }
+      spatial_y_[level]->DerivativeY(src);
+      spatial_y_computed_[level] = true;
+    }
+    return spatial_y_[level];
+  }
+
+  // The integral image is currently only used for object detection, so lazily
+  // initialize it on request.
+  inline const IntegralImage* GetIntegralImage() const {
+    if (integral_image_.get() == NULL) {
+      integral_image_.reset(new IntegralImage(image_));
+    } else if (!integral_image_computed_) {
+      integral_image_->Recompute(image_);
+    }
+    integral_image_computed_ = true;
+    return integral_image_.get();
+  }
+
+  inline const Image<uint8_t>* GetU() const {
+    SCHECK(uv_data_computed_, "UV data not provided!");
+    return u_data_.get();
+  }
+
+  inline const Image<uint8_t>* GetV() const {
+    SCHECK(uv_data_computed_, "UV data not provided!");
+    return v_data_.get();
+  }
+
+ private:
+  void Precompute() {
+    // Create the smoothed pyramids.
+    for (int i = 0; i < kNumPyramidLevels * 2; i += 2) {
+      (void) GetPyramidSqrt2Level(i);
+    }
+    TimeLog("Created smoothed pyramids");
+
+    // Create the smoothed pyramids.
+    for (int i = 1; i < kNumPyramidLevels * 2; i += 2) {
+      (void) GetPyramidSqrt2Level(i);
+    }
+    TimeLog("Created smoothed sqrt pyramids");
+
+    // Create the spatial derivatives for frame 1.
+    for (int i = 0; i < kNumPyramidLevels; ++i) {
+      (void) GetSpatialX(i);
+      (void) GetSpatialY(i);
+    }
+    TimeLog("Created spatial derivatives");
+
+    (void) GetIntegralImage();
+    TimeLog("Got integral image!");
+  }
+
+  const int uv_frame_width_;
+  const int uv_frame_height_;
+
+  int64_t timestamp_;
+
+  Image<uint8_t> image_;
+
+  bool uv_data_computed_;
+  std::unique_ptr<Image<uint8_t> > u_data_;
+  std::unique_ptr<Image<uint8_t> > v_data_;
+
+  mutable bool spatial_x_computed_[kNumPyramidLevels];
+  mutable Image<int32_t>* spatial_x_[kNumPyramidLevels];
+
+  mutable bool spatial_y_computed_[kNumPyramidLevels];
+  mutable Image<int32_t>* spatial_y_[kNumPyramidLevels];
+
+  // Mutable so the lazy initialization can work when this class is const.
+  // Whether or not the integral image has been computed for the current image.
+  mutable bool integral_image_computed_;
+  mutable std::unique_ptr<IntegralImage> integral_image_;
+
+  mutable bool pyramid_sqrt2_computed_[kNumPyramidLevels * 2];
+  mutable Image<uint8_t>* pyramid_sqrt2_[kNumPyramidLevels * 2];
+
+  ImageData(const ImageData&) = delete;
+  void operator=(const ImageData&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_DATA_H_

tensorflow/tools/android/test/jni/object_tracking/image_neon.cc

@@ -0,0 +1,268 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// NEON implementations of Image methods for compatible devices.  Control
+// should never enter this compilation unit on incompatible devices.
+
+#ifdef __ARM_NEON
+
+#include <arm_neon.h>
+#include <stdint.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_utils.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+// This function does the bulk of the work.
+template <>
+void Image<uint8_t>::Downsample2x32ColumnsNeon(const uint8_t* const original,
+                                               const int stride,
+                                               const int orig_x) {
+  // Divide input x offset by 2 to find output offset.
+  const int new_x = orig_x >> 1;
+
+  // Initial offset into top row.
+  const uint8_t* offset = original + orig_x;
+
+  // This points to the leftmost pixel of our 8 horizontally arranged
+  // pixels in the destination data.
+  uint8_t* ptr_dst = (*this)[0] + new_x;
+
+  // Sum along vertical columns.
+  // Process 32x2 input pixels and 16x1 output pixels per iteration.
+  for (int new_y = 0; new_y < height_; ++new_y) {
+    uint16x8_t accum1 = vdupq_n_u16(0);
+    uint16x8_t accum2 = vdupq_n_u16(0);
+
+    // Go top to bottom across the four rows of input pixels that make up
+    // this output row.
+    for (int row_num = 0; row_num < 2; ++row_num) {
+      // First 16 bytes.
+      {
+        // Load 16 bytes of data from current offset.
+        const uint8x16_t curr_data1 = vld1q_u8(offset);
+
+        // Pairwise add and accumulate into accum vectors (16 bit to account
+        // for values above 255).
+        accum1 = vpadalq_u8(accum1, curr_data1);
+      }
+
+      // Second 16 bytes.
+      {
+        // Load 16 bytes of data from current offset.
+        const uint8x16_t curr_data2 = vld1q_u8(offset + 16);
+
+        // Pairwise add and accumulate into accum vectors (16 bit to account
+        // for values above 255).
+        accum2 = vpadalq_u8(accum2, curr_data2);
+      }
+
+      // Move offset down one row.
+      offset += stride;
+    }
+
+    // Divide by 4 (number of input pixels per output
+    // pixel) and narrow data from 16 bits per pixel to 8 bpp.
+    const uint8x8_t tmp_pix1 = vqshrn_n_u16(accum1, 2);
+    const uint8x8_t tmp_pix2 = vqshrn_n_u16(accum2, 2);
+
+    // Concatenate 8x1 pixel strips into 16x1 pixel strip.
+    const uint8x16_t allpixels = vcombine_u8(tmp_pix1, tmp_pix2);
+
+    // Copy all pixels from composite 16x1 vector into output strip.
+    vst1q_u8(ptr_dst, allpixels);
+
+    ptr_dst += stride_;
+  }
+}
+
+// This function does the bulk of the work.
+template <>
+void Image<uint8_t>::Downsample4x32ColumnsNeon(const uint8_t* const original,
+                                               const int stride,
+                                               const int orig_x) {
+  // Divide input x offset by 4 to find output offset.
+  const int new_x = orig_x >> 2;
+
+  // Initial offset into top row.
+  const uint8_t* offset = original + orig_x;
+
+  // This points to the leftmost pixel of our 8 horizontally arranged
+  // pixels in the destination data.
+  uint8_t* ptr_dst = (*this)[0] + new_x;
+
+  // Sum along vertical columns.
+  // Process 32x4 input pixels and 8x1 output pixels per iteration.
+  for (int new_y = 0; new_y < height_; ++new_y) {
+    uint16x8_t accum1 = vdupq_n_u16(0);
+    uint16x8_t accum2 = vdupq_n_u16(0);
+
+    // Go top to bottom across the four rows of input pixels that make up
+    // this output row.
+    for (int row_num = 0; row_num < 4; ++row_num) {
+      // First 16 bytes.
+      {
+        // Load 16 bytes of data from current offset.
+        const uint8x16_t curr_data1 = vld1q_u8(offset);
+
+        // Pairwise add and accumulate into accum vectors (16 bit to account
+        // for values above 255).
+        accum1 = vpadalq_u8(accum1, curr_data1);
+      }
+
+      // Second 16 bytes.
+      {
+        // Load 16 bytes of data from current offset.
+        const uint8x16_t curr_data2 = vld1q_u8(offset + 16);
+
+        // Pairwise add and accumulate into accum vectors (16 bit to account
+        // for values above 255).
+        accum2 = vpadalq_u8(accum2, curr_data2);
+      }
+
+      // Move offset down one row.
+      offset += stride;
+    }
+
+    // Add and widen, then divide by 16 (number of input pixels per output
+    // pixel) and narrow data from 32 bits per pixel to 16 bpp.
+    const uint16x4_t tmp_pix1 = vqshrn_n_u32(vpaddlq_u16(accum1), 4);
+    const uint16x4_t tmp_pix2 = vqshrn_n_u32(vpaddlq_u16(accum2), 4);
+
+    // Combine 4x1 pixel strips into 8x1 pixel strip and narrow from
+    // 16 bits to 8 bits per pixel.
+    const uint8x8_t allpixels = vmovn_u16(vcombine_u16(tmp_pix1, tmp_pix2));
+
+    // Copy all pixels from composite 8x1 vector into output strip.
+    vst1_u8(ptr_dst, allpixels);
+
+    ptr_dst += stride_;
+  }
+}
+
+
+// Hardware accelerated downsampling method for supported devices.
+// Requires that image size be a multiple of 16 pixels in each dimension,
+// and that downsampling be by a factor of 2 or 4.
+template <>
+void Image<uint8_t>::DownsampleAveragedNeon(const uint8_t* const original,
+                                            const int stride,
+                                            const int factor) {
+  // TODO(andrewharp): stride is a bad approximation for the src image's width.
+  // Better to pass that in directly.
+  SCHECK(width_ * factor <= stride, "Uh oh!");
+  const int last_starting_index = width_ * factor - 32;
+
+  // We process 32 input pixels lengthwise at a time.
+  // The output per pass of this loop is an 8 wide by downsampled height tall
+  // pixel strip.
+  int orig_x = 0;
+  for (; orig_x <= last_starting_index; orig_x += 32) {
+    if (factor == 2) {
+      Downsample2x32ColumnsNeon(original, stride, orig_x);
+    } else {
+      Downsample4x32ColumnsNeon(original, stride, orig_x);
+    }
+  }
+
+  // If a last pass is required, push it to the left enough so that it never
+  // goes out of bounds. This will result in some extra computation on devices
+  // whose frame widths are multiples of 16 and not 32.
+  if (orig_x < last_starting_index + 32) {
+    if (factor == 2) {
+      Downsample2x32ColumnsNeon(original, stride, last_starting_index);
+    } else {
+      Downsample4x32ColumnsNeon(original, stride, last_starting_index);
+    }
+  }
+}
+
+
+// Puts the image gradient matrix about a pixel into the 2x2 float array G.
+// vals_x should be an array of the window x gradient values, whose indices
+// can be in any order but are parallel to the vals_y entries.
+// See http://robots.stanford.edu/cs223b04/algo_tracking.pdf for more details.
+void CalculateGNeon(const float* const vals_x, const float* const vals_y,
+                    const int num_vals, float* const G) {
+  const float32_t* const arm_vals_x = (const float32_t*) vals_x;
+  const float32_t* const arm_vals_y = (const float32_t*) vals_y;
+
+  // Running sums.
+  float32x4_t xx = vdupq_n_f32(0.0f);
+  float32x4_t xy = vdupq_n_f32(0.0f);
+  float32x4_t yy = vdupq_n_f32(0.0f);
+
+  // Maximum index we can load 4 consecutive values from.
+  // e.g. if there are 81 values, our last full pass can be from index 77:
+  // 81-4=>77 (77, 78, 79, 80)
+  const int max_i = num_vals - 4;
+
+  // Defined here because we want to keep track of how many values were
+  // processed by NEON, so that we can finish off the remainder the normal
+  // way.
+  int i = 0;
+
+  // Process values 4 at a time, accumulating the sums of
+  // the pixel-wise x*x, x*y, and y*y values.
+  for (; i <= max_i; i += 4) {
+    // Load xs
+    float32x4_t x = vld1q_f32(arm_vals_x + i);
+
+    // Multiply x*x and accumulate.
+    xx = vmlaq_f32(xx, x, x);
+
+    // Load ys
+    float32x4_t y = vld1q_f32(arm_vals_y + i);
+
+    // Multiply x*y and accumulate.
+    xy = vmlaq_f32(xy, x, y);
+
+    // Multiply y*y and accumulate.
+    yy = vmlaq_f32(yy, y, y);
+  }
+
+  static float32_t xx_vals[4];
+  static float32_t xy_vals[4];
+  static float32_t yy_vals[4];
+
+  vst1q_f32(xx_vals, xx);
+  vst1q_f32(xy_vals, xy);
+  vst1q_f32(yy_vals, yy);
+
+  // Accumulated values are store in sets of 4, we have to manually add
+  // the last bits together.
+  for (int j = 0; j < 4; ++j) {
+    G[0] += xx_vals[j];
+    G[1] += xy_vals[j];
+    G[3] += yy_vals[j];
+  }
+
+  // Finishes off last few values (< 4) from above.
+  for (; i < num_vals; ++i) {
+    G[0] += Square(vals_x[i]);
+    G[1] += vals_x[i] * vals_y[i];
+    G[3] += Square(vals_y[i]);
+  }
+
+  // The matrix is symmetric, so this is a given.
+  G[2] = G[1];
+}
+
+}  // namespace tf_tracking
+
+#endif

tensorflow/tools/android/test/jni/object_tracking/image_utils.h

@@ -0,0 +1,297 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_UTILS_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_UTILS_H_
+
+#include <stdint.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+inline void GetUV(const uint8_t* const input, Image<uint8_t>* const u,
+                  Image<uint8_t>* const v) {
+  const uint8_t* pUV = input;
+
+  for (int row = 0; row < u->GetHeight(); ++row) {
+    uint8_t* u_curr = (*u)[row];
+    uint8_t* v_curr = (*v)[row];
+    for (int col = 0; col < u->GetWidth(); ++col) {
+#ifdef __APPLE__
+      *u_curr++ = *pUV++;
+      *v_curr++ = *pUV++;
+#else
+      *v_curr++ = *pUV++;
+      *u_curr++ = *pUV++;
+#endif
+    }
+  }
+}
+
+// Marks every point within a circle of a given radius on the given boolean
+// image true.
+template <typename U>
+inline static void MarkImage(const int x, const int y, const int radius,
+                             Image<U>* const img) {
+  SCHECK(img->ValidPixel(x, y), "Marking invalid pixel in image! %d, %d", x, y);
+
+  // Precomputed for efficiency.
+  const int squared_radius = Square(radius);
+
+  // Mark every row in the circle.
+  for (int d_y = 0; d_y <= radius; ++d_y) {
+    const int squared_y_dist = Square(d_y);
+
+    const int min_y = MAX(y - d_y, 0);
+    const int max_y = MIN(y + d_y, img->height_less_one_);
+
+    // The max d_x of the circle must be strictly greater or equal to
+    // radius - d_y for any positive d_y. Thus, starting from radius - d_y will
+    // reduce the number of iterations required as compared to starting from
+    // either 0 and counting up or radius and counting down.
+    for (int d_x = radius - d_y; d_x <= radius; ++d_x) {
+      // The first time this criteria is met, we know the width of the circle at
+      // this row (without using sqrt).
+      if (squared_y_dist + Square(d_x) >= squared_radius) {
+        const int min_x = MAX(x - d_x, 0);
+        const int max_x = MIN(x + d_x, img->width_less_one_);
+
+        // Mark both above and below the center row.
+        bool* const top_row_start = (*img)[min_y] + min_x;
+        bool* const bottom_row_start = (*img)[max_y] + min_x;
+
+        const int x_width = max_x - min_x + 1;
+        memset(top_row_start, true, sizeof(*top_row_start) * x_width);
+        memset(bottom_row_start, true, sizeof(*bottom_row_start) * x_width);
+
+        // This row is marked, time to move on to the next row.
+        break;
+      }
+    }
+  }
+}
+
+#ifdef __ARM_NEON
+void CalculateGNeon(
+    const float* const vals_x, const float* const vals_y,
+    const int num_vals, float* const G);
+#endif
+
+// Puts the image gradient matrix about a pixel into the 2x2 float array G.
+// vals_x should be an array of the window x gradient values, whose indices
+// can be in any order but are parallel to the vals_y entries.
+// See http://robots.stanford.edu/cs223b04/algo_tracking.pdf for more details.
+inline void CalculateG(const float* const vals_x, const float* const vals_y,
+                       const int num_vals, float* const G) {
+#ifdef __ARM_NEON
+  CalculateGNeon(vals_x, vals_y, num_vals, G);
+  return;
+#endif
+
+  // Non-accelerated version.
+  for (int i = 0; i < num_vals; ++i) {
+    G[0] += Square(vals_x[i]);
+    G[1] += vals_x[i] * vals_y[i];
+    G[3] += Square(vals_y[i]);
+  }
+
+  // The matrix is symmetric, so this is a given.
+  G[2] = G[1];
+}
+
+inline void CalculateGInt16(const int16_t* const vals_x,
+                            const int16_t* const vals_y, const int num_vals,
+                            int* const G) {
+  // Non-accelerated version.
+  for (int i = 0; i < num_vals; ++i) {
+    G[0] += Square(vals_x[i]);
+    G[1] += vals_x[i] * vals_y[i];
+    G[3] += Square(vals_y[i]);
+  }
+
+  // The matrix is symmetric, so this is a given.
+  G[2] = G[1];
+}
+
+
+// Puts the image gradient matrix about a pixel into the 2x2 float array G.
+// Looks up interpolated pixels, then calls above method for implementation.
+inline void CalculateG(const int window_radius, const float center_x,
+                       const float center_y, const Image<int32_t>& I_x,
+                       const Image<int32_t>& I_y, float* const G) {
+  SCHECK(I_x.ValidPixel(center_x, center_y), "Problem in calculateG!");
+
+  // Hardcoded to allow for a max window radius of 5 (9 pixels x 9 pixels).
+  static const int kMaxWindowRadius = 5;
+  SCHECK(window_radius <= kMaxWindowRadius,
+        "Window %d > %d!", window_radius, kMaxWindowRadius);
+
+  // Diameter of window is 2 * radius + 1 for center pixel.
+  static const int kWindowBufferSize =
+      (kMaxWindowRadius * 2 + 1) * (kMaxWindowRadius * 2 + 1);
+
+  // Preallocate buffers statically for efficiency.
+  static int16_t vals_x[kWindowBufferSize];
+  static int16_t vals_y[kWindowBufferSize];
+
+  const int src_left_fixed = RealToFixed1616(center_x - window_radius);
+  const int src_top_fixed = RealToFixed1616(center_y - window_radius);
+
+  int16_t* vals_x_ptr = vals_x;
+  int16_t* vals_y_ptr = vals_y;
+
+  const int window_size = 2 * window_radius + 1;
+  for (int y = 0; y < window_size; ++y) {
+    const int fp_y = src_top_fixed + (y << 16);
+
+    for (int x = 0; x < window_size; ++x) {
+      const int fp_x = src_left_fixed + (x << 16);
+
+      *vals_x_ptr++ = I_x.GetPixelInterpFixed1616(fp_x, fp_y);
+      *vals_y_ptr++ = I_y.GetPixelInterpFixed1616(fp_x, fp_y);
+    }
+  }
+
+  int32_t g_temp[] = {0, 0, 0, 0};
+  CalculateGInt16(vals_x, vals_y, window_size * window_size, g_temp);
+
+  for (int i = 0; i < 4; ++i) {
+    G[i] = g_temp[i];
+  }
+}
+
+inline float ImageCrossCorrelation(const Image<float>& image1,
+                                   const Image<float>& image2,
+                                   const int x_offset, const int y_offset) {
+  SCHECK(image1.GetWidth() == image2.GetWidth() &&
+         image1.GetHeight() == image2.GetHeight(),
+        "Dimension mismatch! %dx%d vs %dx%d",
+        image1.GetWidth(), image1.GetHeight(),
+        image2.GetWidth(), image2.GetHeight());
+
+  const int num_pixels = image1.GetWidth() * image1.GetHeight();
+  const float* data1 = image1.data();
+  const float* data2 = image2.data();
+  return ComputeCrossCorrelation(data1, data2, num_pixels);
+}
+
+// Copies an arbitrary region of an image to another (floating point)
+// image, scaling as it goes using bilinear interpolation.
+inline void CopyArea(const Image<uint8_t>& image,
+                     const BoundingBox& area_to_copy,
+                     Image<float>* const patch_image) {
+  VLOG(2) << "Copying from: " << area_to_copy << std::endl;
+
+  const int patch_width = patch_image->GetWidth();
+  const int patch_height = patch_image->GetHeight();
+
+  const float x_dist_between_samples = patch_width > 0 ?
+      area_to_copy.GetWidth() / (patch_width - 1) : 0;
+
+  const float y_dist_between_samples = patch_height > 0 ?
+      area_to_copy.GetHeight() / (patch_height - 1) : 0;
+
+  for (int y_index = 0; y_index < patch_height; ++y_index) {
+    const float sample_y =
+        y_index * y_dist_between_samples + area_to_copy.top_;
+
+    for (int x_index = 0; x_index < patch_width; ++x_index) {
+      const float sample_x =
+          x_index * x_dist_between_samples + area_to_copy.left_;
+
+      if (image.ValidInterpPixel(sample_x, sample_y)) {
+        // TODO(andrewharp): Do area averaging when downsampling.
+        (*patch_image)[y_index][x_index] =
+            image.GetPixelInterp(sample_x, sample_y);
+      } else {
+        (*patch_image)[y_index][x_index] = -1.0f;
+      }
+    }
+  }
+}
+
+
+// Takes a floating point image and normalizes it in-place.
+//
+// First, negative values will be set to the mean of the non-negative pixels
+// in the image.
+//
+// Then, the resulting will be normalized such that it has mean value of 0.0 and
+// a standard deviation of 1.0.
+inline void NormalizeImage(Image<float>* const image) {
+  const float* const data_ptr = image->data();
+
+  // Copy only the non-negative values to some temp memory.
+  float running_sum = 0.0f;
+  int num_data_gte_zero = 0;
+  {
+    float* const curr_data = (*image)[0];
+    for (int i = 0; i < image->data_size_; ++i) {
+      if (curr_data[i] >= 0.0f) {
+        running_sum += curr_data[i];
+        ++num_data_gte_zero;
+      } else {
+        curr_data[i] = -1.0f;
+      }
+    }
+  }
+
+  // If none of the pixels are valid, just set the entire thing to 0.0f.
+  if (num_data_gte_zero == 0) {
+    image->Clear(0.0f);
+    return;
+  }
+
+  const float corrected_mean = running_sum / num_data_gte_zero;
+
+  float* curr_data = (*image)[0];
+  for (int i = 0; i < image->data_size_; ++i) {
+    const float curr_val = *curr_data;
+    *curr_data++ = curr_val < 0 ? 0 : curr_val - corrected_mean;
+  }
+
+  const float std_dev = ComputeStdDev(data_ptr, image->data_size_, 0.0f);
+
+  if (std_dev > 0.0f) {
+    curr_data = (*image)[0];
+    for (int i = 0; i < image->data_size_; ++i) {
+      *curr_data++ /= std_dev;
+    }
+
+#ifdef SANITY_CHECKS
+    LOGV("corrected_mean: %1.2f  std_dev: %1.2f", corrected_mean, std_dev);
+    const float correlation =
+        ComputeCrossCorrelation(image->data(),
+                                image->data(),
+                                image->data_size_);
+
+    if (std::abs(correlation - 1.0f) > EPSILON) {
+      LOG(ERROR) << "Bad image!" << std::endl;
+      LOG(ERROR) << *image << std::endl;
+    }
+
+    SCHECK(std::abs(correlation - 1.0f) < EPSILON,
+           "Correlation wasn't 1.0f:  %.10f", correlation);
+#endif
+  }
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_IMAGE_UTILS_H_

tensorflow/tools/android/test/jni/object_tracking/integral_image.h

@@ -0,0 +1,188 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_INTEGRAL_IMAGE_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_INTEGRAL_IMAGE_H_
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+typedef uint8_t Code;
+
+class IntegralImage : public Image<uint32_t> {
+ public:
+  explicit IntegralImage(const Image<uint8_t>& image_base)
+      : Image<uint32_t>(image_base.GetWidth(), image_base.GetHeight()) {
+    Recompute(image_base);
+  }
+
+  IntegralImage(const int width, const int height)
+      : Image<uint32_t>(width, height) {}
+
+  void Recompute(const Image<uint8_t>& image_base) {
+    SCHECK(image_base.GetWidth() == GetWidth() &&
+          image_base.GetHeight() == GetHeight(), "Dimensions don't match!");
+
+    // Sum along first row.
+    {
+      int x_sum = 0;
+      for (int x = 0; x < image_base.GetWidth(); ++x) {
+        x_sum += image_base[0][x];
+        (*this)[0][x] = x_sum;
+      }
+    }
+
+    // Sum everything else.
+    for (int y = 1; y < image_base.GetHeight(); ++y) {
+      uint32_t* curr_sum = (*this)[y];
+
+      // Previously summed pointers.
+      const uint32_t* up_one = (*this)[y - 1];
+
+      // Current value pointer.
+      const uint8_t* curr_delta = image_base[y];
+
+      uint32_t row_till_now = 0;
+
+      for (int x = 0; x < GetWidth(); ++x) {
+        // Add the one above and the one to the left.
+        row_till_now += *curr_delta;
+        *curr_sum = *up_one + row_till_now;
+
+        // Scoot everything along.
+        ++curr_sum;
+        ++up_one;
+        ++curr_delta;
+      }
+    }
+
+    SCHECK(VerifyData(image_base), "Images did not match!");
+  }
+
+  bool VerifyData(const Image<uint8_t>& image_base) {
+    for (int y = 0; y < GetHeight(); ++y) {
+      for (int x = 0; x < GetWidth(); ++x) {
+        uint32_t curr_val = (*this)[y][x];
+
+        if (x > 0) {
+          curr_val -= (*this)[y][x - 1];
+        }
+
+        if (y > 0) {
+          curr_val -= (*this)[y - 1][x];
+        }
+
+        if (x > 0 && y > 0) {
+          curr_val += (*this)[y - 1][x - 1];
+        }
+
+        if (curr_val != image_base[y][x]) {
+          LOGE("Mismatch! %d vs %d", curr_val, image_base[y][x]);
+          return false;
+        }
+
+        if (GetRegionSum(x, y, x, y) != curr_val) {
+          LOGE("Mismatch!");
+        }
+      }
+    }
+
+    return true;
+  }
+
+  // Returns the sum of all pixels in the specified region.
+  inline uint32_t GetRegionSum(const int x1, const int y1, const int x2,
+                               const int y2) const {
+    SCHECK(x1 >= 0 && y1 >= 0 &&
+          x2 >= x1 && y2 >= y1 && x2 < GetWidth() && y2 < GetHeight(),
+          "indices out of bounds! %d-%d / %d, %d-%d / %d, ",
+          x1, x2, GetWidth(), y1, y2, GetHeight());
+
+    const uint32_t everything = (*this)[y2][x2];
+
+    uint32_t sum = everything;
+    if (x1 > 0 && y1 > 0) {
+      // Most common case.
+      const uint32_t left = (*this)[y2][x1 - 1];
+      const uint32_t top = (*this)[y1 - 1][x2];
+      const uint32_t top_left = (*this)[y1 - 1][x1 - 1];
+
+      sum = everything - left - top + top_left;
+      SCHECK(sum >= 0, "Both: %d - %d - %d + %d => %d! indices: %d %d %d %d",
+            everything, left, top, top_left, sum, x1, y1, x2, y2);
+    } else if (x1 > 0) {
+      // Flush against top of image.
+      // Subtract out the region to the left only.
+      const uint32_t top = (*this)[y2][x1 - 1];
+      sum = everything - top;
+      SCHECK(sum >= 0, "Top: %d - %d => %d!", everything, top, sum);
+    } else if (y1 > 0) {
+      // Flush against left side of image.
+      // Subtract out the region above only.
+      const uint32_t left = (*this)[y1 - 1][x2];
+      sum = everything - left;
+      SCHECK(sum >= 0, "Left: %d - %d => %d!", everything, left, sum);
+    }
+
+    SCHECK(sum >= 0, "Negative sum!");
+
+    return sum;
+  }
+
+  // Returns the 2bit code associated with this region, which represents
+  // the overall gradient.
+  inline Code GetCode(const BoundingBox& bounding_box) const {
+    return GetCode(bounding_box.left_, bounding_box.top_,
+                   bounding_box.right_, bounding_box.bottom_);
+  }
+
+  inline Code GetCode(const int x1, const int y1,
+                      const int x2, const int y2) const {
+    SCHECK(x1 < x2 && y1 < y2, "Bounds out of order!! TL:%d,%d BR:%d,%d",
+           x1, y1, x2, y2);
+
+    // Gradient computed vertically.
+    const int box_height = (y2 - y1) / 2;
+    const int top_sum = GetRegionSum(x1, y1, x2, y1 + box_height);
+    const int bottom_sum = GetRegionSum(x1, y2 - box_height, x2, y2);
+    const bool vertical_code = top_sum > bottom_sum;
+
+    // Gradient computed horizontally.
+    const int box_width = (x2 - x1) / 2;
+    const int left_sum = GetRegionSum(x1, y1, x1 + box_width, y2);
+    const int right_sum = GetRegionSum(x2 - box_width, y1, x2, y2);
+    const bool horizontal_code = left_sum > right_sum;
+
+    const Code final_code = (vertical_code << 1) | horizontal_code;
+
+    SCHECK(InRange(final_code, static_cast<Code>(0), static_cast<Code>(3)),
+          "Invalid code! %d", final_code);
+
+    // Returns a value 0-3.
+    return final_code;
+  }
+
+ private:
+  IntegralImage(const IntegralImage&) = delete;
+  void operator=(const IntegralImage&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_INTEGRAL_IMAGE_H_

tensorflow/tools/android/test/jni/object_tracking/jni_utils.h

@@ -0,0 +1,64 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_JNI_UTILS_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_JNI_UTILS_H_
+
+#include <jni.h>
+#include <stdint.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+// The JniLongField class is used to access Java fields from native code. This
+// technique of hiding pointers to native objects in opaque Java fields is how
+// the Android hardware libraries work. This reduces the amount of static
+// native methods and makes it easier to manage the lifetime of native objects.
+class JniLongField {
+ public:
+  JniLongField(const char* field_name)
+      : field_name_(field_name), field_ID_(0) {}
+
+  int64_t get(JNIEnv* env, jobject thiz) {
+    if (field_ID_ == 0) {
+      jclass cls = env->GetObjectClass(thiz);
+      CHECK_ALWAYS(cls != 0, "Unable to find class");
+      field_ID_ = env->GetFieldID(cls, field_name_, "J");
+      CHECK_ALWAYS(field_ID_ != 0,
+          "Unable to find field %s. (Check proguard cfg)", field_name_);
+    }
+
+    return env->GetLongField(thiz, field_ID_);
+  }
+
+  void set(JNIEnv* env, jobject thiz, int64_t value) {
+    if (field_ID_ == 0) {
+      jclass cls = env->GetObjectClass(thiz);
+      CHECK_ALWAYS(cls != 0, "Unable to find class");
+      field_ID_ = env->GetFieldID(cls, field_name_, "J");
+      CHECK_ALWAYS(field_ID_ != 0,
+          "Unable to find field %s (Check proguard cfg)", field_name_);
+    }
+
+    env->SetLongField(thiz, field_ID_, value);
+  }
+
+ private:
+  const char* const field_name_;
+
+  // This is just a cache
+  jfieldID field_ID_;
+};
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_JNI_UTILS_H_

tensorflow/tools/android/test/jni/object_tracking/keypoint.h

@@ -0,0 +1,47 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_KEYPOINT_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_KEYPOINT_H_
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+// For keeping track of keypoints.
+struct Keypoint {
+  Keypoint() : pos_(0.0f, 0.0f), score_(0.0f), type_(0) {}
+  Keypoint(const float x, const float y)
+      : pos_(x, y), score_(0.0f), type_(0) {}
+
+  Point2f pos_;
+  float score_;
+  uint8_t type_;
+};
+
+inline std::ostream& operator<<(std::ostream& stream, const Keypoint keypoint) {
+  return stream << "[" << keypoint.pos_ << ", "
+      << keypoint.score_ << ", " << keypoint.type_ << "]";
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_KEYPOINT_H_

tensorflow/tools/android/test/jni/object_tracking/keypoint_detector.cc

@@ -0,0 +1,548 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// Various keypoint detecting functions.
+
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint_detector.h"
+
+#include <float.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+static inline int GetDistSquaredBetween(const int* vec1, const int* vec2) {
+  return Square(vec1[0] - vec2[0]) + Square(vec1[1] - vec2[1]);
+}
+
+void KeypointDetector::ScoreKeypoints(const ImageData& image_data,
+                                      const int num_candidates,
+                                      Keypoint* const candidate_keypoints) {
+  const Image<int>& I_x = *image_data.GetSpatialX(0);
+  const Image<int>& I_y = *image_data.GetSpatialY(0);
+
+  if (config_->detect_skin) {
+    const Image<uint8_t>& u_data = *image_data.GetU();
+    const Image<uint8_t>& v_data = *image_data.GetV();
+
+    static const int reference[] = {111, 155};
+
+    // Score all the keypoints.
+    for (int i = 0; i < num_candidates; ++i) {
+      Keypoint* const keypoint = candidate_keypoints + i;
+
+      const int x_pos = keypoint->pos_.x * 2;
+      const int y_pos = keypoint->pos_.y * 2;
+
+      const int curr_color[] = {u_data[y_pos][x_pos], v_data[y_pos][x_pos]};
+      keypoint->score_ =
+          HarrisFilter(I_x, I_y, keypoint->pos_.x, keypoint->pos_.y) /
+          GetDistSquaredBetween(reference, curr_color);
+    }
+  } else {
+    // Score all the keypoints.
+    for (int i = 0; i < num_candidates; ++i) {
+      Keypoint* const keypoint = candidate_keypoints + i;
+      keypoint->score_ =
+          HarrisFilter(I_x, I_y, keypoint->pos_.x, keypoint->pos_.y);
+    }
+  }
+}
+
+
+inline int KeypointCompare(const void* const a, const void* const b) {
+  return (reinterpret_cast<const Keypoint*>(a)->score_ -
+          reinterpret_cast<const Keypoint*>(b)->score_) <= 0 ? 1 : -1;
+}
+
+
+// Quicksorts detected keypoints by score.
+void KeypointDetector::SortKeypoints(const int num_candidates,
+                                   Keypoint* const candidate_keypoints) const {
+  qsort(candidate_keypoints, num_candidates, sizeof(Keypoint), KeypointCompare);
+
+#ifdef SANITY_CHECKS
+  // Verify that the array got sorted.
+  float last_score = FLT_MAX;
+  for (int i = 0; i < num_candidates; ++i) {
+    const float curr_score = candidate_keypoints[i].score_;
+
+    // Scores should be monotonically increasing.
+    SCHECK(last_score >= curr_score,
+          "Quicksort failure! %d: %.5f > %d: %.5f (%d total)",
+          i - 1, last_score, i, curr_score, num_candidates);
+
+    last_score = curr_score;
+  }
+#endif
+}
+
+
+int KeypointDetector::SelectKeypointsInBox(
+    const BoundingBox& box,
+    const Keypoint* const candidate_keypoints,
+    const int num_candidates,
+    const int max_keypoints,
+    const int num_existing_keypoints,
+    const Keypoint* const existing_keypoints,
+    Keypoint* const final_keypoints) const {
+  if (max_keypoints <= 0) {
+    return 0;
+  }
+
+  // This is the distance within which keypoints may be placed to each other
+  // within this box, roughly based on the box dimensions.
+  const int distance =
+      MAX(1, MIN(box.GetWidth(), box.GetHeight()) * kClosestPercent / 2.0f);
+
+  // First, mark keypoints that already happen to be inside this region. Ignore
+  // keypoints that are outside it, however close they might be.
+  interest_map_->Clear(false);
+  for (int i = 0; i < num_existing_keypoints; ++i) {
+    const Keypoint& candidate = existing_keypoints[i];
+
+    const int x_pos = candidate.pos_.x;
+    const int y_pos = candidate.pos_.y;
+    if (box.Contains(candidate.pos_)) {
+      MarkImage(x_pos, y_pos, distance, interest_map_.get());
+    }
+  }
+
+  // Now, go through and check which keypoints will still fit in the box.
+  int num_keypoints_selected = 0;
+  for (int i = 0; i < num_candidates; ++i) {
+    const Keypoint& candidate = candidate_keypoints[i];
+
+    const int x_pos = candidate.pos_.x;
+    const int y_pos = candidate.pos_.y;
+
+    if (!box.Contains(candidate.pos_) ||
+        !interest_map_->ValidPixel(x_pos, y_pos)) {
+      continue;
+    }
+
+    if (!(*interest_map_)[y_pos][x_pos]) {
+      final_keypoints[num_keypoints_selected++] = candidate;
+      if (num_keypoints_selected >= max_keypoints) {
+        break;
+      }
+      MarkImage(x_pos, y_pos, distance, interest_map_.get());
+    }
+  }
+  return num_keypoints_selected;
+}
+
+
+void KeypointDetector::SelectKeypoints(
+    const std::vector<BoundingBox>& boxes,
+    const Keypoint* const candidate_keypoints,
+    const int num_candidates,
+    FramePair* const curr_change) const {
+  // Now select all the interesting keypoints that fall insider our boxes.
+  curr_change->number_of_keypoints_ = 0;
+  for (std::vector<BoundingBox>::const_iterator iter = boxes.begin();
+      iter != boxes.end(); ++iter) {
+    const BoundingBox bounding_box = *iter;
+
+    // Count up keypoints that have already been selected, and fall within our
+    // box.
+    int num_keypoints_already_in_box = 0;
+    for (int i = 0; i < curr_change->number_of_keypoints_; ++i) {
+      if (bounding_box.Contains(curr_change->frame1_keypoints_[i].pos_)) {
+        ++num_keypoints_already_in_box;
+      }
+    }
+
+    const int max_keypoints_to_find_in_box =
+        MIN(kMaxKeypointsForObject - num_keypoints_already_in_box,
+            kMaxKeypoints - curr_change->number_of_keypoints_);
+
+    const int num_new_keypoints_in_box = SelectKeypointsInBox(
+        bounding_box,
+        candidate_keypoints,
+        num_candidates,
+        max_keypoints_to_find_in_box,
+        curr_change->number_of_keypoints_,
+        curr_change->frame1_keypoints_,
+        curr_change->frame1_keypoints_ + curr_change->number_of_keypoints_);
+
+    curr_change->number_of_keypoints_ += num_new_keypoints_in_box;
+
+    LOGV("Selected %d keypoints!", curr_change->number_of_keypoints_);
+  }
+}
+
+
+// Walks along the given circle checking for pixels above or below the center.
+// Returns a score, or 0 if the keypoint did not pass the criteria.
+//
+// Parameters:
+//  circle_perimeter: the circumference in pixels of the circle.
+//  threshold: the minimum number of contiguous pixels that must be above or
+//             below the center value.
+//  center_ptr: the location of the center pixel in memory
+//  offsets: the relative offsets from the center pixel of the edge pixels.
+inline int TestCircle(const int circle_perimeter, const int threshold,
+                      const uint8_t* const center_ptr, const int* offsets) {
+  // Get the actual value of the center pixel for easier reference later on.
+  const int center_value = static_cast<int>(*center_ptr);
+
+  // Number of total pixels to check.  Have to wrap around some in case
+  // the contiguous section is split by the array edges.
+  const int num_total = circle_perimeter + threshold - 1;
+
+  int num_above = 0;
+  int above_diff = 0;
+
+  int num_below = 0;
+  int below_diff = 0;
+
+  // Used to tell when this is definitely not going to meet the threshold so we
+  // can early abort.
+  int minimum_by_now = threshold - num_total + 1;
+
+  // Go through every pixel along the perimeter of the circle, and then around
+  // again a little bit.
+  for (int i = 0; i < num_total; ++i) {
+    // This should be faster than mod.
+    const int perim_index = i < circle_perimeter ? i : i - circle_perimeter;
+
+    // This gets the value of the current pixel along the perimeter by using
+    // a precomputed offset.
+    const int curr_value =
+        static_cast<int>(center_ptr[offsets[perim_index]]);
+
+    const int difference = curr_value - center_value;
+
+    if (difference > kFastDiffAmount) {
+      above_diff += difference;
+      ++num_above;
+
+      num_below = 0;
+      below_diff = 0;
+
+      if (num_above >= threshold) {
+        return above_diff;
+      }
+    } else if (difference < -kFastDiffAmount) {
+      below_diff += difference;
+      ++num_below;
+
+      num_above = 0;
+      above_diff = 0;
+
+      if (num_below >= threshold) {
+        return below_diff;
+      }
+    } else {
+      num_above = 0;
+      num_below = 0;
+      above_diff = 0;
+      below_diff = 0;
+    }
+
+    // See if there's any chance of making the threshold.
+    if (MAX(num_above, num_below) < minimum_by_now) {
+      // Didn't pass.
+      return 0;
+    }
+    ++minimum_by_now;
+  }
+
+  // Didn't pass.
+  return 0;
+}
+
+
+// Returns a score in the range [0.0, positive infinity) which represents the
+// relative likelihood of a point being a corner.
+float KeypointDetector::HarrisFilter(const Image<int32_t>& I_x,
+                                     const Image<int32_t>& I_y, const float x,
+                                     const float y) const {
+  if (I_x.ValidInterpPixel(x - kHarrisWindowSize, y - kHarrisWindowSize) &&
+      I_x.ValidInterpPixel(x + kHarrisWindowSize, y + kHarrisWindowSize)) {
+    // Image gradient matrix.
+    float G[] = { 0, 0, 0, 0 };
+    CalculateG(kHarrisWindowSize, x, y, I_x, I_y, G);
+
+    const float dx = G[0];
+    const float dy = G[3];
+    const float dxy = G[1];
+
+    // Harris-Nobel corner score.
+    return (dx * dy - Square(dxy)) / (dx + dy + FLT_MIN);
+  }
+
+  return 0.0f;
+}
+
+
+int KeypointDetector::AddExtraCandidatesForBoxes(
+    const std::vector<BoundingBox>& boxes,
+    const int max_num_keypoints,
+    Keypoint* const keypoints) const {
+  int num_keypoints_added = 0;
+
+  for (std::vector<BoundingBox>::const_iterator iter = boxes.begin();
+      iter != boxes.end(); ++iter) {
+    const BoundingBox box = *iter;
+
+    for (int i = 0; i < kNumToAddAsCandidates; ++i) {
+      for (int j = 0; j < kNumToAddAsCandidates; ++j) {
+        if (num_keypoints_added >= max_num_keypoints) {
+          LOGW("Hit cap of %d for temporary keypoints!", max_num_keypoints);
+          return num_keypoints_added;
+        }
+
+        Keypoint& curr_keypoint = keypoints[num_keypoints_added++];
+        curr_keypoint.pos_ = Point2f(
+            box.left_ + box.GetWidth() * (i + 0.5f) / kNumToAddAsCandidates,
+            box.top_ + box.GetHeight() * (j + 0.5f) / kNumToAddAsCandidates);
+        curr_keypoint.type_ = KEYPOINT_TYPE_INTEREST;
+      }
+    }
+  }
+
+  return num_keypoints_added;
+}
+
+
+void KeypointDetector::FindKeypoints(const ImageData& image_data,
+                                   const std::vector<BoundingBox>& rois,
+                                   const FramePair& prev_change,
+                                   FramePair* const curr_change) {
+  // Copy keypoints from second frame of last pass to temp keypoints of this
+  // pass.
+  int number_of_tmp_keypoints = CopyKeypoints(prev_change, tmp_keypoints_);
+
+  const int max_num_fast = kMaxTempKeypoints - number_of_tmp_keypoints;
+  number_of_tmp_keypoints +=
+      FindFastKeypoints(image_data, max_num_fast,
+                       tmp_keypoints_ + number_of_tmp_keypoints);
+
+  TimeLog("Found FAST keypoints");
+
+  if (number_of_tmp_keypoints >= kMaxTempKeypoints) {
+    LOGW("Hit cap of %d for temporary keypoints (FAST)! %d keypoints",
+         kMaxTempKeypoints, number_of_tmp_keypoints);
+  }
+
+  if (kAddArbitraryKeypoints) {
+    // Add some for each object prior to scoring.
+    const int max_num_box_keypoints =
+        kMaxTempKeypoints - number_of_tmp_keypoints;
+    number_of_tmp_keypoints +=
+        AddExtraCandidatesForBoxes(rois, max_num_box_keypoints,
+                                   tmp_keypoints_ + number_of_tmp_keypoints);
+    TimeLog("Added box keypoints");
+
+    if (number_of_tmp_keypoints >= kMaxTempKeypoints) {
+      LOGW("Hit cap of %d for temporary keypoints (boxes)! %d keypoints",
+           kMaxTempKeypoints, number_of_tmp_keypoints);
+    }
+  }
+
+  // Score them...
+  LOGV("Scoring %d keypoints!", number_of_tmp_keypoints);
+  ScoreKeypoints(image_data, number_of_tmp_keypoints, tmp_keypoints_);
+  TimeLog("Scored keypoints");
+
+  // Now pare it down a bit.
+  SortKeypoints(number_of_tmp_keypoints, tmp_keypoints_);
+  TimeLog("Sorted keypoints");
+
+  LOGV("%d keypoints to select from!", number_of_tmp_keypoints);
+
+  SelectKeypoints(rois, tmp_keypoints_, number_of_tmp_keypoints, curr_change);
+  TimeLog("Selected keypoints");
+
+  LOGV("Picked %d (%d max) final keypoints out of %d potential.",
+       curr_change->number_of_keypoints_,
+       kMaxKeypoints, number_of_tmp_keypoints);
+}
+
+
+int KeypointDetector::CopyKeypoints(const FramePair& prev_change,
+                                  Keypoint* const new_keypoints) {
+  int number_of_keypoints = 0;
+
+  // Caching values from last pass, just copy and compact.
+  for (int i = 0; i < prev_change.number_of_keypoints_; ++i) {
+    if (prev_change.optical_flow_found_keypoint_[i]) {
+      new_keypoints[number_of_keypoints] =
+          prev_change.frame2_keypoints_[i];
+
+      new_keypoints[number_of_keypoints].score_ =
+          prev_change.frame1_keypoints_[i].score_;
+
+      ++number_of_keypoints;
+    }
+  }
+
+  TimeLog("Copied keypoints");
+  return number_of_keypoints;
+}
+
+
+// FAST keypoint detector.
+int KeypointDetector::FindFastKeypoints(const Image<uint8_t>& frame,
+                                        const int quadrant,
+                                        const int downsample_factor,
+                                        const int max_num_keypoints,
+                                        Keypoint* const keypoints) {
+  /*
+   // Reference for a circle of diameter 7.
+   const int circle[] = {0, 0, 1, 1, 1, 0, 0,
+                         0, 1, 0, 0, 0, 1, 0,
+                         1, 0, 0, 0, 0, 0, 1,
+                         1, 0, 0, 0, 0, 0, 1,
+                         1, 0, 0, 0, 0, 0, 1,
+                         0, 1, 0, 0, 0, 1, 0,
+                         0, 0, 1, 1, 1, 0, 0};
+   const int circle_offset[] =
+       {2, 3, 4, 8, 12, 14, 20, 21, 27, 28, 34, 36, 40, 44, 45, 46};
+   */
+
+  // Quick test of compass directions.  Any length 16 circle with a break of up
+  // to 4 pixels will have at least 3 of these 4 pixels active.
+  static const int short_circle_perimeter = 4;
+  static const int short_threshold = 3;
+  static const int short_circle_x[] = { -3,  0, +3,  0 };
+  static const int short_circle_y[] = {  0, -3,  0, +3 };
+
+  // Precompute image offsets.
+  int short_offsets[short_circle_perimeter];
+  for (int i = 0; i < short_circle_perimeter; ++i) {
+    short_offsets[i] = short_circle_x[i] + short_circle_y[i] * frame.GetWidth();
+  }
+
+  // Large circle values.
+  static const int full_circle_perimeter = 16;
+  static const int full_threshold = 12;
+  static const int full_circle_x[] =
+      { -1,  0, +1, +2, +3, +3, +3, +2, +1, +0, -1, -2, -3, -3, -3, -2 };
+  static const int full_circle_y[] =
+      { -3, -3, -3, -2, -1,  0, +1, +2, +3, +3, +3, +2, +1, +0, -1, -2 };
+
+  // Precompute image offsets.
+  int full_offsets[full_circle_perimeter];
+  for (int i = 0; i < full_circle_perimeter; ++i) {
+    full_offsets[i] = full_circle_x[i] + full_circle_y[i] * frame.GetWidth();
+  }
+
+  const int scratch_stride = frame.stride();
+
+  keypoint_scratch_->Clear(0);
+
+  // Set up the bounds on the region to test based on the passed-in quadrant.
+  const int quadrant_width = (frame.GetWidth() / 2) - kFastBorderBuffer;
+  const int quadrant_height = (frame.GetHeight() / 2) - kFastBorderBuffer;
+  const int start_x =
+      kFastBorderBuffer + ((quadrant % 2 == 0) ? 0 : quadrant_width);
+  const int start_y =
+      kFastBorderBuffer + ((quadrant < 2) ? 0 : quadrant_height);
+  const int end_x = start_x + quadrant_width;
+  const int end_y = start_y + quadrant_height;
+
+  // Loop through once to find FAST keypoint clumps.
+  for (int img_y = start_y; img_y < end_y; ++img_y) {
+    const uint8_t* curr_pixel_ptr = frame[img_y] + start_x;
+
+    for (int img_x = start_x; img_x < end_x; ++img_x) {
+      // Only insert it if it meets the quick minimum requirements test.
+      if (TestCircle(short_circle_perimeter, short_threshold,
+                     curr_pixel_ptr, short_offsets) != 0) {
+        // Longer test for actual keypoint score..
+        const int fast_score = TestCircle(full_circle_perimeter,
+                                          full_threshold,
+                                          curr_pixel_ptr,
+                                          full_offsets);
+
+        // Non-zero score means the keypoint was found.
+        if (fast_score != 0) {
+          uint8_t* const center_ptr = (*keypoint_scratch_)[img_y] + img_x;
+
+          // Increase the keypoint count on this pixel and the pixels in all
+          // 4 cardinal directions.
+          *center_ptr += 5;
+          *(center_ptr - 1) += 1;
+          *(center_ptr + 1) += 1;
+          *(center_ptr - scratch_stride) += 1;
+          *(center_ptr + scratch_stride) += 1;
+        }
+      }
+
+      ++curr_pixel_ptr;
+    }  // x
+  }  // y
+
+  TimeLog("Found FAST keypoints.");
+
+  int num_keypoints = 0;
+  // Loop through again and Harris filter pixels in the center of clumps.
+  // We can shrink the window by 1 pixel on every side.
+  for (int img_y = start_y + 1; img_y < end_y - 1; ++img_y) {
+    const uint8_t* curr_pixel_ptr = (*keypoint_scratch_)[img_y] + start_x;
+
+    for (int img_x = start_x + 1; img_x < end_x - 1; ++img_x) {
+      if (*curr_pixel_ptr >= kMinNumConnectedForFastKeypoint) {
+       Keypoint* const keypoint = keypoints + num_keypoints;
+        keypoint->pos_ = Point2f(
+            img_x * downsample_factor, img_y * downsample_factor);
+        keypoint->score_ = 0;
+        keypoint->type_ = KEYPOINT_TYPE_FAST;
+
+        ++num_keypoints;
+        if (num_keypoints >= max_num_keypoints) {
+          return num_keypoints;
+        }
+      }
+
+      ++curr_pixel_ptr;
+    }  // x
+  }  // y
+
+  TimeLog("Picked FAST keypoints.");
+
+  return num_keypoints;
+}
+
+int KeypointDetector::FindFastKeypoints(const ImageData& image_data,
+                                        const int max_num_keypoints,
+                                        Keypoint* const keypoints) {
+  int downsample_factor = 1;
+  int num_found = 0;
+
+  // TODO(andrewharp): Get this working for multiple image scales.
+  for (int i = 0; i < 1; ++i) {
+    const Image<uint8_t>& frame = *image_data.GetPyramidSqrt2Level(i);
+    num_found += FindFastKeypoints(
+        frame, fast_quadrant_,
+        downsample_factor, max_num_keypoints, keypoints + num_found);
+    downsample_factor *= 2;
+  }
+
+  // Increment the current quadrant.
+  fast_quadrant_ = (fast_quadrant_ + 1) % 4;
+
+  return num_found;
+}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/keypoint_detector.h

@@ -0,0 +1,129 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_KEYPOINT_DETECTOR_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_KEYPOINT_DETECTOR_H_
+
+#include <stdint.h>
+
+#include <vector>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_data.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/optical_flow.h"
+
+namespace tf_tracking {
+
+struct Keypoint;
+
+class KeypointDetector {
+ public:
+  explicit KeypointDetector(const KeypointDetectorConfig* const config)
+      : config_(config),
+        keypoint_scratch_(new Image<uint8_t>(config_->image_size)),
+        interest_map_(new Image<bool>(config_->image_size)),
+        fast_quadrant_(0) {
+    interest_map_->Clear(false);
+  }
+
+  ~KeypointDetector() {}
+
+  // Finds a new set of keypoints for the current frame, picked from the current
+  // set of keypoints and also from a set discovered via a keypoint detector.
+  // Special attention is applied to make sure that keypoints are distributed
+  // within the supplied ROIs.
+  void FindKeypoints(const ImageData& image_data,
+                     const std::vector<BoundingBox>& rois,
+                     const FramePair& prev_change,
+                     FramePair* const curr_change);
+
+ private:
+  // Compute the corneriness of a point in the image.
+  float HarrisFilter(const Image<int32_t>& I_x, const Image<int32_t>& I_y,
+                     const float x, const float y) const;
+
+  // Adds a grid of candidate keypoints to the given box, up to
+  // max_num_keypoints or kNumToAddAsCandidates^2, whichever is lower.
+  int AddExtraCandidatesForBoxes(
+      const std::vector<BoundingBox>& boxes,
+      const int max_num_keypoints,
+      Keypoint* const keypoints) const;
+
+  // Scan the frame for potential keypoints using the FAST keypoint detector.
+  // Quadrant is an argument 0-3 which refers to the quadrant of the image in
+  // which to detect keypoints.
+  int FindFastKeypoints(const Image<uint8_t>& frame, const int quadrant,
+                        const int downsample_factor,
+                        const int max_num_keypoints, Keypoint* const keypoints);
+
+  int FindFastKeypoints(const ImageData& image_data,
+                        const int max_num_keypoints,
+                        Keypoint* const keypoints);
+
+  // Score a bunch of candidate keypoints.  Assigns the scores to the input
+  // candidate_keypoints array entries.
+  void ScoreKeypoints(const ImageData& image_data,
+                      const int num_candidates,
+                      Keypoint* const candidate_keypoints);
+
+  void SortKeypoints(const int num_candidates,
+                    Keypoint* const candidate_keypoints) const;
+
+  // Selects a set of keypoints falling within the supplied box such that the
+  // most highly rated keypoints are picked first, and so that none of them are
+  // too close together.
+  int SelectKeypointsInBox(
+      const BoundingBox& box,
+      const Keypoint* const candidate_keypoints,
+      const int num_candidates,
+      const int max_keypoints,
+      const int num_existing_keypoints,
+      const Keypoint* const existing_keypoints,
+      Keypoint* const final_keypoints) const;
+
+  // Selects from the supplied sorted keypoint pool a set of keypoints that will
+  // best cover the given set of boxes, such that each box is covered at a
+  // resolution proportional to its size.
+  void SelectKeypoints(
+      const std::vector<BoundingBox>& boxes,
+      const Keypoint* const candidate_keypoints,
+      const int num_candidates,
+      FramePair* const frame_change) const;
+
+  // Copies and compacts the found keypoints in the second frame of prev_change
+  // into the array at new_keypoints.
+  static int CopyKeypoints(const FramePair& prev_change,
+                          Keypoint* const new_keypoints);
+
+  const KeypointDetectorConfig* const config_;
+
+  // Scratch memory for keypoint candidacy detection and non-max suppression.
+  std::unique_ptr<Image<uint8_t> > keypoint_scratch_;
+
+  // Regions of the image to pay special attention to.
+  std::unique_ptr<Image<bool> > interest_map_;
+
+  // The current quadrant of the image to detect FAST keypoints in.
+  // Keypoint detection is staggered for performance reasons. Every four frames
+  // a full scan of the frame will have been performed.
+  int fast_quadrant_;
+
+  Keypoint tmp_keypoints_[kMaxTempKeypoints];
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_KEYPOINT_DETECTOR_H_

tensorflow/tools/android/test/jni/object_tracking/logging.cc

@@ -0,0 +1,140 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+
+#ifdef STANDALONE_DEMO_LIB
+
+#include <android/log.h>
+#include <stdlib.h>
+#include <time.h>
+#include <iostream>
+#include <sstream>
+
+LogMessage::LogMessage(const char* fname, int line, int severity)
+    : fname_(fname), line_(line), severity_(severity) {}
+
+void LogMessage::GenerateLogMessage() {
+  int android_log_level;
+  switch (severity_) {
+    case INFO:
+      android_log_level = ANDROID_LOG_INFO;
+      break;
+    case WARNING:
+      android_log_level = ANDROID_LOG_WARN;
+      break;
+    case ERROR:
+      android_log_level = ANDROID_LOG_ERROR;
+      break;
+    case FATAL:
+      android_log_level = ANDROID_LOG_FATAL;
+      break;
+    default:
+      if (severity_ < INFO) {
+        android_log_level = ANDROID_LOG_VERBOSE;
+      } else {
+        android_log_level = ANDROID_LOG_ERROR;
+      }
+      break;
+  }
+
+  std::stringstream ss;
+  const char* const partial_name = strrchr(fname_, '/');
+  ss << (partial_name != nullptr ? partial_name + 1 : fname_) << ":" << line_
+     << " " << str();
+  __android_log_write(android_log_level, "native", ss.str().c_str());
+
+  // Also log to stderr (for standalone Android apps).
+  std::cerr << "native : " << ss.str() << std::endl;
+
+  // Android logging at level FATAL does not terminate execution, so abort()
+  // is still required to stop the program.
+  if (severity_ == FATAL) {
+    abort();
+  }
+}
+
+namespace {
+
+// Parse log level (int64) from environment variable (char*)
+int64_t LogLevelStrToInt(const char* tf_env_var_val) {
+  if (tf_env_var_val == nullptr) {
+    return 0;
+  }
+
+  // Ideally we would use env_var / safe_strto64, but it is
+  // hard to use here without pulling in a lot of dependencies,
+  // so we use std:istringstream instead
+  std::string min_log_level(tf_env_var_val);
+  std::istringstream ss(min_log_level);
+  int64_t level;
+  if (!(ss >> level)) {
+    // Invalid vlog level setting, set level to default (0)
+    level = 0;
+  }
+
+  return level;
+}
+
+int64_t MinLogLevelFromEnv() {
+  const char* tf_env_var_val = getenv("TF_CPP_MIN_LOG_LEVEL");
+  return LogLevelStrToInt(tf_env_var_val);
+}
+
+int64_t MinVLogLevelFromEnv() {
+  const char* tf_env_var_val = getenv("TF_CPP_MIN_VLOG_LEVEL");
+  return LogLevelStrToInt(tf_env_var_val);
+}
+
+}  // namespace
+
+LogMessage::~LogMessage() {
+  // Read the min log level once during the first call to logging.
+  static int64_t min_log_level = MinLogLevelFromEnv();
+  if (TF_PREDICT_TRUE(severity_ >= min_log_level)) GenerateLogMessage();
+}
+
+int64_t LogMessage::MinVLogLevel() {
+  static const int64_t min_vlog_level = MinVLogLevelFromEnv();
+  return min_vlog_level;
+}
+
+LogMessageFatal::LogMessageFatal(const char* file, int line)
+    : LogMessage(file, line, ANDROID_LOG_FATAL) {}
+LogMessageFatal::~LogMessageFatal() {
+  // abort() ensures we don't return (we promised we would not via
+  // ATTRIBUTE_NORETURN).
+  GenerateLogMessage();
+  abort();
+}
+
+void LogString(const char* fname, int line, int severity,
+               const std::string& message) {
+  LogMessage(fname, line, severity) << message;
+}
+
+void LogPrintF(const int severity, const char* format, ...) {
+  char message[1024];
+  va_list argptr;
+  va_start(argptr, format);
+  vsnprintf(message, 1024, format, argptr);
+  va_end(argptr);
+  __android_log_write(severity, "native", message);
+
+  // Also log to stderr (for standalone Android apps).
+  std::cerr << "native : " << message << std::endl;
+}
+
+#endif

tensorflow/tools/android/test/jni/object_tracking/logging.h

@@ -0,0 +1,121 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_LOGGING_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_LOGGING_H_
+
+#include <android/log.h>
+#include <string.h>
+#include <ostream>
+#include <sstream>
+#include <string>
+
+// Allow this library to be built without depending on TensorFlow by
+// defining STANDALONE_DEMO_LIB. Otherwise TensorFlow headers will be
+// used.
+#ifdef STANDALONE_DEMO_LIB
+
+// A macro to disallow the copy constructor and operator= functions
+// This is usually placed in the private: declarations for a class.
+#define TF_DISALLOW_COPY_AND_ASSIGN(TypeName) \
+  TypeName(const TypeName&) = delete;         \
+  void operator=(const TypeName&) = delete
+
+#if defined(__GNUC__)
+#define TF_PREDICT_FALSE(x) (__builtin_expect(x, 0))
+#define TF_PREDICT_TRUE(x) (__builtin_expect(!!(x), 1))
+#else
+#define TF_PREDICT_FALSE(x) (x)
+#define TF_PREDICT_TRUE(x) (x)
+#endif
+
+// Log levels equivalent to those defined by
+// third_party/tensorflow/core/platform/logging.h
+const int INFO = 0;            // base_logging::INFO;
+const int WARNING = 1;         // base_logging::WARNING;
+const int ERROR = 2;           // base_logging::ERROR;
+const int FATAL = 3;           // base_logging::FATAL;
+const int NUM_SEVERITIES = 4;  // base_logging::NUM_SEVERITIES;
+
+class LogMessage : public std::basic_ostringstream<char> {
+ public:
+  LogMessage(const char* fname, int line, int severity);
+  ~LogMessage();
+
+  // Returns the minimum log level for VLOG statements.
+  // E.g., if MinVLogLevel() is 2, then VLOG(2) statements will produce output,
+  // but VLOG(3) will not. Defaults to 0.
+  static int64_t MinVLogLevel();
+
+ protected:
+  void GenerateLogMessage();
+
+ private:
+  const char* fname_;
+  int line_;
+  int severity_;
+};
+
+// LogMessageFatal ensures the process will exit in failure after
+// logging this message.
+class LogMessageFatal : public LogMessage {
+ public:
+  LogMessageFatal(const char* file, int line);
+  ~LogMessageFatal();
+};
+
+#define _TF_LOG_INFO \
+  ::tensorflow::internal::LogMessage(__FILE__, __LINE__, tensorflow::INFO)
+#define _TF_LOG_WARNING \
+  ::tensorflow::internal::LogMessage(__FILE__, __LINE__, tensorflow::WARNING)
+#define _TF_LOG_ERROR \
+  ::tensorflow::internal::LogMessage(__FILE__, __LINE__, tensorflow::ERROR)
+#define _TF_LOG_FATAL \
+  ::tensorflow::internal::LogMessageFatal(__FILE__, __LINE__)
+
+#define _TF_LOG_QFATAL _TF_LOG_FATAL
+
+#define LOG(severity) _TF_LOG_##severity
+
+#define VLOG_IS_ON(lvl) ((lvl) <= LogMessage::MinVLogLevel())
+
+#define VLOG(lvl)                        \
+  if (TF_PREDICT_FALSE(VLOG_IS_ON(lvl))) \
+  LogMessage(__FILE__, __LINE__, ANDROID_LOG_INFO)
+
+void LogPrintF(const int severity, const char* format, ...);
+
+// Support for printf style logging.
+#define LOGV(...)
+#define LOGD(...)
+#define LOGI(...) LogPrintF(ANDROID_LOG_INFO, __VA_ARGS__);
+#define LOGW(...) LogPrintF(ANDROID_LOG_INFO, __VA_ARGS__);
+#define LOGE(...) LogPrintF(ANDROID_LOG_ERROR, __VA_ARGS__);
+
+#else
+
+#include "tensorflow/core/lib/strings/stringprintf.h"
+#include "tensorflow/core/platform/logging.h"
+
+// Support for printf style logging.
+#define LOGV(...)
+#define LOGD(...)
+#define LOGI(...) LOG(INFO) << tensorflow::strings::Printf(__VA_ARGS__);
+#define LOGW(...) LOG(INFO) << tensorflow::strings::Printf(__VA_ARGS__);
+#define LOGE(...) LOG(INFO) << tensorflow::strings::Printf(__VA_ARGS__);
+
+#endif
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_LOGGING_H_

tensorflow/tools/android/test/jni/object_tracking/object_detector.cc

@@ -0,0 +1,27 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// NOTE: no native object detectors are currently provided or used by the code
+// in this directory. This class remains mainly for historical reasons.
+// Detection in the TF demo is done through TensorFlowMultiBoxDetector.java.
+
+#include "tensorflow/tools/android/test/jni/object_tracking/object_detector.h"
+
+namespace tf_tracking {
+
+// This is here so that the vtable gets created properly.
+ObjectDetectorBase::~ObjectDetectorBase() {}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/object_detector.h

@@ -0,0 +1,232 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// NOTE: no native object detectors are currently provided or used by the code
+// in this directory. This class remains mainly for historical reasons.
+// Detection in the TF demo is done through TensorFlowMultiBoxDetector.java.
+
+// Defines the ObjectDetector class that is the main interface for detecting
+// ObjectModelBases in frames.
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_DETECTOR_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_DETECTOR_H_
+
+#include <float.h>
+
+#include <map>
+#include <memory>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/integral_image.h"
+#ifdef __RENDER_OPENGL__
+#include "tensorflow/tools/android/test/jni/object_tracking/sprite.h"
+#endif
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_data.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/object_model.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+// Adds BoundingSquares to a vector such that the first square added is centered
+// in the position given and of square_size, and the remaining squares are added
+// concentrentically, scaling down by scale_factor until the minimum threshold
+// size is passed.
+// Squares that do not fall completely within image_bounds will not be added.
+static inline void FillWithSquares(
+    const BoundingBox& image_bounds,
+    const BoundingBox& position,
+    const float starting_square_size,
+    const float smallest_square_size,
+    const float scale_factor,
+    std::vector<BoundingSquare>* const squares) {
+  BoundingSquare descriptor_area =
+      GetCenteredSquare(position, starting_square_size);
+
+  SCHECK(scale_factor < 1.0f, "Scale factor too large at %.2f!", scale_factor);
+
+  // Use a do/while loop to ensure that at least one descriptor is created.
+  do {
+    if (image_bounds.Contains(descriptor_area.ToBoundingBox())) {
+      squares->push_back(descriptor_area);
+    }
+    descriptor_area.Scale(scale_factor);
+  } while (descriptor_area.size_ >= smallest_square_size - EPSILON);
+  LOGV("Created %zu squares starting from size %.2f to min size %.2f "
+       "using scale factor: %.2f",
+       squares->size(), starting_square_size, smallest_square_size,
+       scale_factor);
+}
+
+
+// Represents a potential detection of a specific ObjectExemplar and Descriptor
+// at a specific position in the image.
+class Detection {
+ public:
+  explicit Detection(const ObjectModelBase* const object_model,
+                     const MatchScore match_score,
+                     const BoundingBox& bounding_box)
+      : object_model_(object_model),
+        match_score_(match_score),
+        bounding_box_(bounding_box) {}
+
+  Detection(const Detection& other)
+      : object_model_(other.object_model_),
+        match_score_(other.match_score_),
+        bounding_box_(other.bounding_box_) {}
+
+  virtual ~Detection() {}
+
+  inline BoundingBox GetObjectBoundingBox() const {
+    return bounding_box_;
+  }
+
+  inline MatchScore GetMatchScore() const {
+    return match_score_;
+  }
+
+  inline const ObjectModelBase* GetObjectModel() const {
+    return object_model_;
+  }
+
+  inline bool Intersects(const Detection& other) {
+    // Check if any of the four axes separates us, there must be at least one.
+    return bounding_box_.Intersects(other.bounding_box_);
+  }
+
+  struct Comp {
+    inline bool operator()(const Detection& a, const Detection& b) const {
+      return a.match_score_ > b.match_score_;
+    }
+  };
+
+  // TODO(andrewharp): add accessors to update these instead.
+  const ObjectModelBase* object_model_;
+  MatchScore match_score_;
+  BoundingBox bounding_box_;
+};
+
+inline std::ostream& operator<<(std::ostream& stream,
+                                const Detection& detection) {
+  const BoundingBox actual_area = detection.GetObjectBoundingBox();
+  stream << actual_area;
+  return stream;
+}
+
+class ObjectDetectorBase {
+ public:
+  explicit ObjectDetectorBase(const ObjectDetectorConfig* const config)
+      : config_(config),
+        image_data_(NULL) {}
+
+  virtual ~ObjectDetectorBase();
+
+  // Sets the current image data. All calls to ObjectDetector other than
+  // FillDescriptors use the image data last set.
+  inline void SetImageData(const ImageData* const image_data) {
+    image_data_ = image_data;
+  }
+
+  // Main entry point into the detection algorithm.
+  // Scans the frame for candidates, tweaks them, and fills in the
+  // given std::vector of Detection objects with acceptable matches.
+  virtual void Detect(const std::vector<BoundingSquare>& positions,
+                      std::vector<Detection>* const detections) const = 0;
+
+  virtual ObjectModelBase* CreateObjectModel(const std::string& name) = 0;
+
+  virtual void DeleteObjectModel(const std::string& name) = 0;
+
+  virtual void GetObjectModels(
+      std::vector<const ObjectModelBase*>* models) const = 0;
+
+  // Creates a new ObjectExemplar from the given position in the context of
+  // the last frame passed to NextFrame.
+  // Will return null in the case that there's no room for a descriptor to be
+  // created in the example area, or the example area is not completely
+  // contained within the frame.
+  virtual void UpdateModel(const Image<uint8_t>& base_image,
+                           const IntegralImage& integral_image,
+                           const BoundingBox& bounding_box, const bool locked,
+                           ObjectModelBase* model) const = 0;
+
+  virtual void Draw() const = 0;
+
+  virtual bool AllowSpontaneousDetections() = 0;
+
+ protected:
+  const std::unique_ptr<const ObjectDetectorConfig> config_;
+
+  // The latest frame data, upon which all detections will be performed.
+  // Not owned by this object, just provided for reference by ObjectTracker
+  // via SetImageData().
+  const ImageData* image_data_;
+
+ private:
+  ObjectDetectorBase(const ObjectDetectorBase&) = delete;
+  void operator=(const ObjectDetectorBase&) = delete;
+};
+
+template <typename ModelType>
+class ObjectDetector : public ObjectDetectorBase {
+ public:
+  explicit ObjectDetector(const ObjectDetectorConfig* const config)
+      : ObjectDetectorBase(config) {}
+
+  virtual ~ObjectDetector() {
+    typename std::map<std::string, ModelType*>::const_iterator it =
+        object_models_.begin();
+    for (; it != object_models_.end(); ++it) {
+      ModelType* model = it->second;
+      delete model;
+    }
+  }
+
+  virtual void DeleteObjectModel(const std::string& name) {
+    ModelType* model = object_models_[name];
+    CHECK_ALWAYS(model != NULL, "Model was null!");
+    object_models_.erase(name);
+    SAFE_DELETE(model);
+  }
+
+  virtual void GetObjectModels(
+      std::vector<const ObjectModelBase*>* models) const {
+    typename std::map<std::string, ModelType*>::const_iterator it =
+        object_models_.begin();
+    for (; it != object_models_.end(); ++it) {
+      models->push_back(it->second);
+    }
+  }
+
+  virtual bool AllowSpontaneousDetections() {
+    return false;
+  }
+
+ protected:
+  std::map<std::string, ModelType*> object_models_;
+
+ private:
+  ObjectDetector(const ObjectDetector&) = delete;
+  void operator=(const ObjectDetector&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_DETECTOR_H_

tensorflow/tools/android/test/jni/object_tracking/object_model.h

@@ -0,0 +1,103 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// NOTE: no native object detectors are currently provided or used by the code
+// in this directory. This class remains mainly for historical reasons.
+// Detection in the TF demo is done through TensorFlowMultiBoxDetector.java.
+
+// Contains ObjectModelBase declaration.
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_MODEL_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_MODEL_H_
+
+#ifdef __RENDER_OPENGL__
+#include <GLES/gl.h>
+#include <GLES/glext.h>
+#endif
+
+#include <vector>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/integral_image.h"
+#ifdef __RENDER_OPENGL__
+#include "tensorflow/tools/android/test/jni/object_tracking/sprite.h"
+#endif
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_data.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+// The ObjectModelBase class represents all the known appearance information for
+// an object. It is not a specific instance of the object in the world,
+// but just the general appearance information that enables detection. An
+// ObjectModelBase can be reused across multiple-instances of TrackedObjects.
+class ObjectModelBase {
+ public:
+  ObjectModelBase(const std::string& name) : name_(name) {}
+
+  virtual ~ObjectModelBase() {}
+
+  // Called when the next step in an ongoing track occurs.
+  virtual void TrackStep(const BoundingBox& position,
+                         const Image<uint8_t>& image,
+                         const IntegralImage& integral_image,
+                         const bool authoritative) {}
+
+  // Called when an object track is lost.
+  virtual void TrackLost() {}
+
+  // Called when an object track is confirmed as legitimate.
+  virtual void TrackConfirmed() {}
+
+  virtual float GetMaxCorrelation(const Image<float>& patch_image) const = 0;
+
+  virtual MatchScore GetMatchScore(
+      const BoundingBox& position, const ImageData& image_data) const = 0;
+
+  virtual void Draw(float* const depth) const = 0;
+
+  inline const std::string& GetName() const {
+    return name_;
+  }
+
+ protected:
+  const std::string name_;
+
+ private:
+  ObjectModelBase(const ObjectModelBase&) = delete;
+  void operator=(const ObjectModelBase&) = delete;
+};
+
+template <typename DetectorType>
+class ObjectModel : public ObjectModelBase {
+ public:
+  ObjectModel<DetectorType>(const DetectorType* const detector,
+                            const std::string& name)
+      : ObjectModelBase(name), detector_(detector) {}
+
+ protected:
+  const DetectorType* const detector_;
+
+  ObjectModel<DetectorType>(const ObjectModel<DetectorType>&) = delete;
+  void operator=(const ObjectModel<DetectorType>&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_MODEL_H_

tensorflow/tools/android/test/jni/object_tracking/object_tracker.cc

@@ -0,0 +1,684 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifdef __RENDER_OPENGL__
+#include <GLES/gl.h>
+#include <GLES/glext.h>
+#endif
+
+#include <cinttypes>
+#include <map>
+#include <string>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/flow_cache.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/integral_image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint_detector.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/object_detector.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/object_tracker.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/optical_flow.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+ObjectTracker::ObjectTracker(const TrackerConfig* const config,
+                             ObjectDetectorBase* const detector)
+    : config_(config),
+      frame_width_(config->image_size.width),
+      frame_height_(config->image_size.height),
+      curr_time_(0),
+      num_frames_(0),
+      flow_cache_(&config->flow_config),
+      keypoint_detector_(&config->keypoint_detector_config),
+      curr_num_frame_pairs_(0),
+      first_frame_index_(0),
+      frame1_(new ImageData(frame_width_, frame_height_)),
+      frame2_(new ImageData(frame_width_, frame_height_)),
+      detector_(detector),
+      num_detected_(0) {
+  for (int i = 0; i < kNumFrames; ++i) {
+    frame_pairs_[i].Init(-1, -1);
+  }
+}
+
+
+ObjectTracker::~ObjectTracker() {
+  for (TrackedObjectMap::iterator iter = objects_.begin();
+       iter != objects_.end(); iter++) {
+    TrackedObject* object = iter->second;
+    SAFE_DELETE(object);
+  }
+}
+
+
+// Finds the correspondences for all the points in the current pair of frames.
+// Stores the results in the given FramePair.
+void ObjectTracker::FindCorrespondences(FramePair* const frame_pair) const {
+  // Keypoints aren't found until they're found.
+  memset(frame_pair->optical_flow_found_keypoint_, false,
+         sizeof(*frame_pair->optical_flow_found_keypoint_) * kMaxKeypoints);
+  TimeLog("Cleared old found keypoints");
+
+  int num_keypoints_found = 0;
+
+  // For every keypoint...
+  for (int i_feat = 0; i_feat < frame_pair->number_of_keypoints_; ++i_feat) {
+    Keypoint* const keypoint1 = frame_pair->frame1_keypoints_ + i_feat;
+    Keypoint* const keypoint2 = frame_pair->frame2_keypoints_ + i_feat;
+
+    if (flow_cache_.FindNewPositionOfPoint(
+        keypoint1->pos_.x, keypoint1->pos_.y,
+        &keypoint2->pos_.x, &keypoint2->pos_.y)) {
+      frame_pair->optical_flow_found_keypoint_[i_feat] = true;
+      ++num_keypoints_found;
+    }
+  }
+
+  TimeLog("Found correspondences");
+
+  LOGV("Found %d of %d keypoint correspondences",
+       num_keypoints_found, frame_pair->number_of_keypoints_);
+}
+
+void ObjectTracker::NextFrame(const uint8_t* const new_frame,
+                              const uint8_t* const uv_frame,
+                              const int64_t timestamp,
+                              const float* const alignment_matrix_2x3) {
+  IncrementFrameIndex();
+  LOGV("Received frame %d", num_frames_);
+
+  FramePair* const curr_change = frame_pairs_ + GetNthIndexFromEnd(0);
+  curr_change->Init(curr_time_, timestamp);
+
+  CHECK_ALWAYS(curr_time_ < timestamp,
+               "Timestamp must monotonically increase! Went from %" PRId64
+               " to %" PRId64 " on frame %d.",
+               curr_time_, timestamp, num_frames_);
+
+  curr_time_ = timestamp;
+
+  // Swap the frames.
+  frame1_.swap(frame2_);
+
+  frame2_->SetData(new_frame, uv_frame, frame_width_, timestamp, 1);
+
+  if (detector_.get() != NULL) {
+    detector_->SetImageData(frame2_.get());
+  }
+
+  flow_cache_.NextFrame(frame2_.get(), alignment_matrix_2x3);
+
+  if (num_frames_ == 1) {
+    // This must be the first frame, so abort.
+    return;
+  }
+
+  if (config_->always_track || objects_.size() > 0) {
+    LOGV("Tracking %zu targets", objects_.size());
+    ComputeKeypoints(true);
+    TimeLog("Keypoints computed!");
+
+    FindCorrespondences(curr_change);
+    TimeLog("Flow computed!");
+
+    TrackObjects();
+  }
+  TimeLog("Targets tracked!");
+
+  if (detector_.get() != NULL && num_frames_ % kDetectEveryNFrames == 0) {
+    DetectTargets();
+  }
+  TimeLog("Detected objects.");
+}
+
+TrackedObject* ObjectTracker::MaybeAddObject(
+    const std::string& id, const Image<uint8_t>& source_image,
+    const BoundingBox& bounding_box, const ObjectModelBase* object_model) {
+  // Train the detector if this is a new object.
+  if (objects_.find(id) != objects_.end()) {
+    return objects_[id];
+  }
+
+  // Need to get a non-const version of the model, or create a new one if it
+  // wasn't given.
+  ObjectModelBase* model = NULL;
+  if (detector_ != NULL) {
+    // If a detector is registered, then this new object must have a model.
+    CHECK_ALWAYS(object_model != NULL, "No model given!");
+    model = detector_->CreateObjectModel(object_model->GetName());
+  }
+  TrackedObject* const object =
+      new TrackedObject(id, source_image, bounding_box, model);
+
+  objects_[id] = object;
+  return object;
+}
+
+void ObjectTracker::RegisterNewObjectWithAppearance(
+    const std::string& id, const uint8_t* const new_frame,
+    const BoundingBox& bounding_box) {
+  ObjectModelBase* object_model = NULL;
+
+  Image<uint8_t> image(frame_width_, frame_height_);
+  image.FromArray(new_frame, frame_width_, 1);
+
+  if (detector_ != NULL) {
+    object_model = detector_->CreateObjectModel(id);
+    CHECK_ALWAYS(object_model != NULL, "Null object model!");
+
+    const IntegralImage integral_image(image);
+    object_model->TrackStep(bounding_box, image, integral_image, true);
+  }
+
+  // Create an object at this position.
+  CHECK_ALWAYS(!HaveObject(id), "Already have this object!");
+  if (objects_.find(id) == objects_.end()) {
+    TrackedObject* const object =
+        MaybeAddObject(id, image, bounding_box, object_model);
+    CHECK_ALWAYS(object != NULL, "Object not created!");
+  }
+}
+
+void ObjectTracker::SetPreviousPositionOfObject(const std::string& id,
+                                                const BoundingBox& bounding_box,
+                                                const int64_t timestamp) {
+  CHECK_ALWAYS(timestamp > 0, "Timestamp too low! %" PRId64, timestamp);
+  CHECK_ALWAYS(timestamp <= curr_time_,
+               "Timestamp too great! %" PRId64 " vs %" PRId64, timestamp,
+               curr_time_);
+
+  TrackedObject* const object = GetObject(id);
+
+  // Track this bounding box from the past to the current time.
+  const BoundingBox current_position = TrackBox(bounding_box, timestamp);
+
+  object->UpdatePosition(current_position, curr_time_, *frame2_, false);
+
+  VLOG(2) << "Set tracked position for " << id << " to " << bounding_box
+          << std::endl;
+}
+
+
+void ObjectTracker::SetCurrentPositionOfObject(
+    const std::string& id, const BoundingBox& bounding_box) {
+  SetPreviousPositionOfObject(id, bounding_box, curr_time_);
+}
+
+
+void ObjectTracker::ForgetTarget(const std::string& id) {
+  LOGV("Forgetting object %s", id.c_str());
+  TrackedObject* const object = GetObject(id);
+  delete object;
+  objects_.erase(id);
+
+  if (detector_ != NULL) {
+    detector_->DeleteObjectModel(id);
+  }
+}
+
+int ObjectTracker::GetKeypointsPacked(uint16_t* const out_data,
+                                      const float scale) const {
+  const FramePair& change = frame_pairs_[GetNthIndexFromEnd(0)];
+  uint16_t* curr_data = out_data;
+  int num_keypoints = 0;
+
+  for (int i = 0; i < change.number_of_keypoints_; ++i) {
+    if (change.optical_flow_found_keypoint_[i]) {
+      ++num_keypoints;
+      const Point2f& point1 = change.frame1_keypoints_[i].pos_;
+      *curr_data++ = RealToFixed115(point1.x * scale);
+      *curr_data++ = RealToFixed115(point1.y * scale);
+
+      const Point2f& point2 = change.frame2_keypoints_[i].pos_;
+      *curr_data++ = RealToFixed115(point2.x * scale);
+      *curr_data++ = RealToFixed115(point2.y * scale);
+    }
+  }
+
+  return num_keypoints;
+}
+
+
+int ObjectTracker::GetKeypoints(const bool only_found,
+                                float* const out_data) const {
+  int curr_keypoint = 0;
+  const FramePair& change = frame_pairs_[GetNthIndexFromEnd(0)];
+
+  for (int i = 0; i < change.number_of_keypoints_; ++i) {
+    if (!only_found || change.optical_flow_found_keypoint_[i]) {
+      const int base = curr_keypoint * kKeypointStep;
+      out_data[base + 0] = change.frame1_keypoints_[i].pos_.x;
+      out_data[base + 1] = change.frame1_keypoints_[i].pos_.y;
+
+      out_data[base + 2] =
+          change.optical_flow_found_keypoint_[i] ? 1.0f : -1.0f;
+      out_data[base + 3] = change.frame2_keypoints_[i].pos_.x;
+      out_data[base + 4] = change.frame2_keypoints_[i].pos_.y;
+
+      out_data[base + 5] = change.frame1_keypoints_[i].score_;
+      out_data[base + 6] = change.frame1_keypoints_[i].type_;
+      ++curr_keypoint;
+    }
+  }
+
+  LOGV("Got %d keypoints.", curr_keypoint);
+
+  return curr_keypoint;
+}
+
+
+BoundingBox ObjectTracker::TrackBox(const BoundingBox& region,
+                                    const FramePair& frame_pair) const {
+  float translation_x;
+  float translation_y;
+
+  float scale_x;
+  float scale_y;
+
+  BoundingBox tracked_box(region);
+  frame_pair.AdjustBox(
+      tracked_box, &translation_x, &translation_y, &scale_x, &scale_y);
+
+  tracked_box.Shift(Point2f(translation_x, translation_y));
+
+  if (scale_x > 0 && scale_y > 0) {
+    tracked_box.Scale(scale_x, scale_y);
+  }
+  return tracked_box;
+}
+
+BoundingBox ObjectTracker::TrackBox(const BoundingBox& region,
+                                    const int64_t timestamp) const {
+  CHECK_ALWAYS(timestamp > 0, "Timestamp too low! %" PRId64, timestamp);
+  CHECK_ALWAYS(timestamp <= curr_time_, "Timestamp is in the future!");
+
+  // Anything that ended before the requested timestamp is of no concern to us.
+  bool found_it = false;
+  int num_frames_back = -1;
+  for (int i = 0; i < curr_num_frame_pairs_; ++i) {
+    const FramePair& frame_pair =
+        frame_pairs_[GetNthIndexFromEnd(i)];
+
+    if (frame_pair.end_time_ <= timestamp) {
+      num_frames_back = i - 1;
+
+      if (num_frames_back > 0) {
+        LOGV("Went %d out of %d frames before finding frame. (index: %d)",
+             num_frames_back, curr_num_frame_pairs_, GetNthIndexFromEnd(i));
+      }
+
+      found_it = true;
+      break;
+    }
+  }
+
+  if (!found_it) {
+    LOGW("History did not go back far enough! %" PRId64 " vs %" PRId64,
+         frame_pairs_[GetNthIndexFromEnd(0)].end_time_ -
+             frame_pairs_[GetNthIndexFromStart(0)].end_time_,
+         frame_pairs_[GetNthIndexFromEnd(0)].end_time_ - timestamp);
+  }
+
+  // Loop over all the frames in the queue, tracking the accumulated delta
+  // of the point from frame to frame.  It's possible the point could
+  // go out of frame, but keep tracking as best we can, using points near
+  // the edge of the screen where it went out of bounds.
+  BoundingBox tracked_box(region);
+  for (int i = num_frames_back; i >= 0; --i) {
+    const FramePair& frame_pair = frame_pairs_[GetNthIndexFromEnd(i)];
+    SCHECK(frame_pair.end_time_ >= timestamp, "Frame timestamp was too early!");
+    tracked_box = TrackBox(tracked_box, frame_pair);
+  }
+  return tracked_box;
+}
+
+
+// Converts a row-major 3x3 2d transformation matrix to a column-major 4x4
+// 3d transformation matrix.
+inline void Convert3x3To4x4(
+    const float* const in_matrix, float* const out_matrix) {
+  // X
+  out_matrix[0] = in_matrix[0];
+  out_matrix[1] = in_matrix[3];
+  out_matrix[2] = 0.0f;
+  out_matrix[3] = 0.0f;
+
+  // Y
+  out_matrix[4] = in_matrix[1];
+  out_matrix[5] = in_matrix[4];
+  out_matrix[6] = 0.0f;
+  out_matrix[7] = 0.0f;
+
+  // Z
+  out_matrix[8] = 0.0f;
+  out_matrix[9] = 0.0f;
+  out_matrix[10] = 1.0f;
+  out_matrix[11] = 0.0f;
+
+  // Translation
+  out_matrix[12] = in_matrix[2];
+  out_matrix[13] = in_matrix[5];
+  out_matrix[14] = 0.0f;
+  out_matrix[15] = 1.0f;
+}
+
+
+void ObjectTracker::Draw(const int canvas_width, const int canvas_height,
+                         const float* const frame_to_canvas) const {
+#ifdef __RENDER_OPENGL__
+  glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+
+  glMatrixMode(GL_PROJECTION);
+  glLoadIdentity();
+
+  glOrthof(0.0f, canvas_width, 0.0f, canvas_height, 0.0f, 1.0f);
+
+  // To make Y go the right direction (0 at top of frame).
+  glScalef(1.0f, -1.0f, 1.0f);
+  glTranslatef(0.0f, -canvas_height, 0.0f);
+
+  glMatrixMode(GL_MODELVIEW);
+  glLoadIdentity();
+
+  glPushMatrix();
+
+  // Apply the frame to canvas transformation.
+  static GLfloat transformation[16];
+  Convert3x3To4x4(frame_to_canvas, transformation);
+  glMultMatrixf(transformation);
+
+  // Draw tracked object bounding boxes.
+  for (TrackedObjectMap::const_iterator iter = objects_.begin();
+    iter != objects_.end(); ++iter) {
+    TrackedObject* tracked_object = iter->second;
+    tracked_object->Draw();
+  }
+
+  static const bool kRenderDebugPyramid = false;
+  if (kRenderDebugPyramid) {
+    glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
+    for (int i = 0; i < kNumPyramidLevels * 2; ++i) {
+      Sprite(*frame1_->GetPyramidSqrt2Level(i)).Draw();
+    }
+  }
+
+  static const bool kRenderDebugDerivative = false;
+  if (kRenderDebugDerivative) {
+    glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
+    for (int i = 0; i < kNumPyramidLevels; ++i) {
+      const Image<int32_t>& dx = *frame1_->GetSpatialX(i);
+      Image<uint8_t> render_image(dx.GetWidth(), dx.GetHeight());
+      for (int y = 0; y < dx.GetHeight(); ++y) {
+        const int32_t* dx_ptr = dx[y];
+        uint8_t* dst_ptr = render_image[y];
+        for (int x = 0; x < dx.GetWidth(); ++x) {
+          *dst_ptr++ = Clip(-(*dx_ptr++), 0, 255);
+        }
+      }
+
+      Sprite(render_image).Draw();
+    }
+  }
+
+  if (detector_ != NULL) {
+    glDisable(GL_CULL_FACE);
+    detector_->Draw();
+  }
+  glPopMatrix();
+#endif
+}
+
+static void AddQuadrants(const BoundingBox& box,
+                         std::vector<BoundingBox>* boxes) {
+  const Point2f center = box.GetCenter();
+
+  float x1 = box.left_;
+  float x2 = center.x;
+  float x3 = box.right_;
+
+  float y1 = box.top_;
+  float y2 = center.y;
+  float y3 = box.bottom_;
+
+  // Upper left.
+  boxes->push_back(BoundingBox(x1, y1, x2, y2));
+
+  // Upper right.
+  boxes->push_back(BoundingBox(x2, y1, x3, y2));
+
+  // Bottom left.
+  boxes->push_back(BoundingBox(x1, y2, x2, y3));
+
+  // Bottom right.
+  boxes->push_back(BoundingBox(x2, y2, x3, y3));
+
+  // Whole thing.
+  boxes->push_back(box);
+}
+
+void ObjectTracker::ComputeKeypoints(const bool cached_ok) {
+  const FramePair& prev_change = frame_pairs_[GetNthIndexFromEnd(1)];
+  FramePair* const curr_change = &frame_pairs_[GetNthIndexFromEnd(0)];
+
+  std::vector<BoundingBox> boxes;
+
+  for (TrackedObjectMap::iterator object_iter = objects_.begin();
+       object_iter != objects_.end(); ++object_iter) {
+    BoundingBox box = object_iter->second->GetPosition();
+    box.Scale(config_->object_box_scale_factor_for_features,
+              config_->object_box_scale_factor_for_features);
+    AddQuadrants(box, &boxes);
+  }
+
+  AddQuadrants(frame1_->GetImage()->GetContainingBox(), &boxes);
+
+  keypoint_detector_.FindKeypoints(*frame1_, boxes, prev_change, curr_change);
+}
+
+
+// Given a vector of detections and a model, simply returns the Detection for
+// that model with the highest correlation.
+bool ObjectTracker::GetBestObjectForDetection(
+    const Detection& detection, TrackedObject** match) const {
+  TrackedObject* best_match = NULL;
+  float best_overlap = -FLT_MAX;
+
+  LOGV("Looking for matches in %zu objects!", objects_.size());
+  for (TrackedObjectMap::const_iterator object_iter = objects_.begin();
+      object_iter != objects_.end(); ++object_iter) {
+    TrackedObject* const tracked_object = object_iter->second;
+
+    const float overlap = tracked_object->GetPosition().PascalScore(
+        detection.GetObjectBoundingBox());
+
+    if (!detector_->AllowSpontaneousDetections() &&
+        (detection.GetObjectModel() != tracked_object->GetModel())) {
+      if (overlap > 0.0f) {
+        return false;
+      }
+      continue;
+    }
+
+    const float jump_distance =
+        (tracked_object->GetPosition().GetCenter() -
+         detection.GetObjectBoundingBox().GetCenter()).LengthSquared();
+
+    const float allowed_distance =
+        tracked_object->GetAllowableDistanceSquared();
+
+    LOGV("Distance: %.2f, Allowed distance %.2f, Overlap: %.2f",
+         jump_distance, allowed_distance, overlap);
+
+    // TODO(andrewharp): No need to do this verification twice, eliminate
+    // one of the score checks (the other being in OnDetection).
+    if (jump_distance < allowed_distance &&
+        overlap > best_overlap &&
+        tracked_object->GetMatchScore() + kMatchScoreBuffer <
+        detection.GetMatchScore()) {
+      best_match = tracked_object;
+      best_overlap = overlap;
+    } else if (overlap > 0.0f) {
+      return false;
+    }
+  }
+
+  *match = best_match;
+  return true;
+}
+
+
+void ObjectTracker::ProcessDetections(
+    std::vector<Detection>* const detections) {
+  LOGV("Initial detection done, iterating over %zu detections now.",
+       detections->size());
+
+  const bool spontaneous_detections_allowed =
+      detector_->AllowSpontaneousDetections();
+  for (std::vector<Detection>::const_iterator it = detections->begin();
+      it != detections->end(); ++it) {
+    const Detection& detection = *it;
+    SCHECK(frame2_->GetImage()->Contains(detection.GetObjectBoundingBox()),
+          "Frame does not contain bounding box!");
+
+    TrackedObject* best_match = NULL;
+
+    const bool no_collisions =
+        GetBestObjectForDetection(detection, &best_match);
+
+    // Need to get a non-const version of the model, or create a new one if it
+    // wasn't given.
+    ObjectModelBase* model =
+        const_cast<ObjectModelBase*>(detection.GetObjectModel());
+
+    if (best_match != NULL) {
+      if (model != best_match->GetModel()) {
+        CHECK_ALWAYS(detector_->AllowSpontaneousDetections(),
+            "Model for object changed but spontaneous detections not allowed!");
+      }
+      best_match->OnDetection(model,
+                              detection.GetObjectBoundingBox(),
+                              detection.GetMatchScore(),
+                              curr_time_, *frame2_);
+    } else if (no_collisions && spontaneous_detections_allowed) {
+      if (detection.GetMatchScore() > kMinimumMatchScore) {
+        LOGV("No match, adding it!");
+        const ObjectModelBase* model = detection.GetObjectModel();
+        std::ostringstream ss;
+        // TODO(andrewharp): Generate this in a more general fashion.
+        ss << "hand_" << num_detected_++;
+        std::string object_name = ss.str();
+        MaybeAddObject(object_name, *frame2_->GetImage(),
+                       detection.GetObjectBoundingBox(), model);
+      }
+    }
+  }
+}
+
+
+void ObjectTracker::DetectTargets() {
+  // Detect all object model types that we're currently tracking.
+  std::vector<const ObjectModelBase*> object_models;
+  detector_->GetObjectModels(&object_models);
+  if (object_models.size() == 0) {
+    LOGV("No objects to search for, aborting.");
+    return;
+  }
+
+  LOGV("Trying to detect %zu models", object_models.size());
+
+  LOGV("Creating test vector!");
+  std::vector<BoundingSquare> positions;
+
+  for (TrackedObjectMap::iterator object_iter = objects_.begin();
+      object_iter != objects_.end(); ++object_iter) {
+    TrackedObject* const tracked_object = object_iter->second;
+
+#if DEBUG_PREDATOR
+  positions.push_back(GetCenteredSquare(
+      frame2_->GetImage()->GetContainingBox(), 32.0f));
+#else
+    const BoundingBox& position = tracked_object->GetPosition();
+
+    const float square_size = MAX(
+        kScanMinSquareSize / (kLastKnownPositionScaleFactor *
+        kLastKnownPositionScaleFactor),
+        MIN(position.GetWidth(),
+        position.GetHeight())) / kLastKnownPositionScaleFactor;
+
+    FillWithSquares(frame2_->GetImage()->GetContainingBox(),
+                    tracked_object->GetPosition(),
+                    square_size,
+                    kScanMinSquareSize,
+                    kLastKnownPositionScaleFactor,
+                    &positions);
+  }
+#endif
+
+  LOGV("Created test vector!");
+
+  std::vector<Detection> detections;
+  LOGV("Detecting!");
+  detector_->Detect(positions, &detections);
+  LOGV("Found %zu detections", detections.size());
+
+  TimeLog("Finished detection.");
+
+  ProcessDetections(&detections);
+
+  TimeLog("iterated over detections");
+
+  LOGV("Done detecting!");
+}
+
+
+void ObjectTracker::TrackObjects() {
+  // TODO(andrewharp): Correlation should be allowed to remove objects too.
+  const bool automatic_removal_allowed = detector_.get() != NULL ?
+      detector_->AllowSpontaneousDetections() : false;
+
+  LOGV("Tracking %zu objects!", objects_.size());
+  std::vector<std::string> dead_objects;
+  for (TrackedObjectMap::iterator iter = objects_.begin();
+       iter != objects_.end(); iter++) {
+    TrackedObject* object = iter->second;
+    const BoundingBox tracked_position = TrackBox(
+        object->GetPosition(), frame_pairs_[GetNthIndexFromEnd(0)]);
+    object->UpdatePosition(tracked_position, curr_time_, *frame2_, false);
+
+    if (automatic_removal_allowed &&
+        object->GetNumConsecutiveFramesBelowThreshold() >
+        kMaxNumDetectionFailures * 5) {
+      dead_objects.push_back(iter->first);
+    }
+  }
+
+  if (detector_ != NULL && automatic_removal_allowed) {
+    for (std::vector<std::string>::iterator iter = dead_objects.begin();
+         iter != dead_objects.end(); iter++) {
+      LOGE("Removing object! %s", iter->c_str());
+      ForgetTarget(*iter);
+    }
+  }
+  TimeLog("Tracked all objects.");
+
+  LOGV("%zu objects tracked!", objects_.size());
+}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/object_tracker.h

@@ -0,0 +1,270 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_TRACKER_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_TRACKER_H_
+
+#include <map>
+#include <string>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/flow_cache.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/integral_image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint_detector.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/object_model.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/optical_flow.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/tracked_object.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+typedef std::map<const std::string, TrackedObject*> TrackedObjectMap;
+
+inline std::ostream& operator<<(std::ostream& stream,
+                                const TrackedObjectMap& map) {
+  for (TrackedObjectMap::const_iterator iter = map.begin();
+      iter != map.end(); ++iter) {
+    const TrackedObject& tracked_object = *iter->second;
+    const std::string& key = iter->first;
+    stream << key << ": " << tracked_object;
+  }
+  return stream;
+}
+
+
+// ObjectTracker is the highest-level class in the tracking/detection framework.
+// It handles basic image processing, keypoint detection, keypoint tracking,
+// object tracking, and object detection/relocalization.
+class ObjectTracker {
+ public:
+  ObjectTracker(const TrackerConfig* const config,
+                ObjectDetectorBase* const detector);
+  virtual ~ObjectTracker();
+
+  virtual void NextFrame(const uint8_t* const new_frame,
+                         const int64_t timestamp,
+                         const float* const alignment_matrix_2x3) {
+    NextFrame(new_frame, NULL, timestamp, alignment_matrix_2x3);
+  }
+
+  // Called upon the arrival of a new frame of raw data.
+  // Does all image processing, keypoint detection, and object
+  // tracking/detection for registered objects.
+  // Argument alignment_matrix_2x3 is a 2x3 matrix (stored row-wise) that
+  // represents the main transformation that has happened between the last
+  // and the current frame.
+  // Argument align_level is the pyramid level (where 0 == finest) that
+  // the matrix is valid for.
+  virtual void NextFrame(const uint8_t* const new_frame,
+                         const uint8_t* const uv_frame, const int64_t timestamp,
+                         const float* const alignment_matrix_2x3);
+
+  virtual void RegisterNewObjectWithAppearance(const std::string& id,
+                                               const uint8_t* const new_frame,
+                                               const BoundingBox& bounding_box);
+
+  // Updates the position of a tracked object, given that it was known to be at
+  // a certain position at some point in the past.
+  virtual void SetPreviousPositionOfObject(const std::string& id,
+                                           const BoundingBox& bounding_box,
+                                           const int64_t timestamp);
+
+  // Sets the current position of the object in the most recent frame provided.
+  virtual void SetCurrentPositionOfObject(const std::string& id,
+                                          const BoundingBox& bounding_box);
+
+  // Tells the ObjectTracker to stop tracking a target.
+  void ForgetTarget(const std::string& id);
+
+  // Fills the given out_data buffer with the latest detected keypoint
+  // correspondences, first scaled by scale_factor (to adjust for downsampling
+  // that may have occurred elsewhere), then packed in a fixed-point format.
+  int GetKeypointsPacked(uint16_t* const out_data,
+                         const float scale_factor) const;
+
+  // Copy the keypoint arrays after computeFlow is called.
+  // out_data should be at least kMaxKeypoints * kKeypointStep long.
+  // Currently, its format is [x1 y1 found x2 y2 score] repeated N times,
+  // where N is the number of keypoints tracked.  N is returned as the result.
+  int GetKeypoints(const bool only_found, float* const out_data) const;
+
+  // Returns the current position of a box, given that it was at a certain
+  // position at the given time.
+  BoundingBox TrackBox(const BoundingBox& region,
+                       const int64_t timestamp) const;
+
+  // Returns the number of frames that have been passed to NextFrame().
+  inline int GetNumFrames() const {
+    return num_frames_;
+  }
+
+  inline bool HaveObject(const std::string& id) const {
+    return objects_.find(id) != objects_.end();
+  }
+
+  // Returns the TrackedObject associated with the given id.
+  inline const TrackedObject* GetObject(const std::string& id) const {
+    TrackedObjectMap::const_iterator iter = objects_.find(id);
+    CHECK_ALWAYS(iter != objects_.end(),
+                 "Unknown object key! \"%s\"", id.c_str());
+    TrackedObject* const object = iter->second;
+    return object;
+  }
+
+  // Returns the TrackedObject associated with the given id.
+  inline TrackedObject* GetObject(const std::string& id) {
+    TrackedObjectMap::iterator iter = objects_.find(id);
+    CHECK_ALWAYS(iter != objects_.end(),
+                 "Unknown object key! \"%s\"", id.c_str());
+    TrackedObject* const object = iter->second;
+    return object;
+  }
+
+  bool IsObjectVisible(const std::string& id) const {
+    SCHECK(HaveObject(id), "Don't have this object.");
+
+    const TrackedObject* object = GetObject(id);
+    return object->IsVisible();
+  }
+
+  virtual void Draw(const int canvas_width, const int canvas_height,
+                    const float* const frame_to_canvas) const;
+
+ protected:
+  // Creates a new tracked object at the given position.
+  // If an object model is provided, then that model will be associated with the
+  // object. If not, a new model may be created from the appearance at the
+  // initial position and registered with the object detector.
+  virtual TrackedObject* MaybeAddObject(const std::string& id,
+                                        const Image<uint8_t>& image,
+                                        const BoundingBox& bounding_box,
+                                        const ObjectModelBase* object_model);
+
+  // Find the keypoints in the frame before the current frame.
+  // If only one frame exists, keypoints will be found in that frame.
+  void ComputeKeypoints(const bool cached_ok = false);
+
+  // Finds the correspondences for all the points in the current pair of frames.
+  // Stores the results in the given FramePair.
+  void FindCorrespondences(FramePair* const curr_change) const;
+
+  inline int GetNthIndexFromEnd(const int offset) const {
+    return GetNthIndexFromStart(curr_num_frame_pairs_ - 1 - offset);
+  }
+
+  BoundingBox TrackBox(const BoundingBox& region,
+                       const FramePair& frame_pair) const;
+
+  inline void IncrementFrameIndex() {
+    // Move the current framechange index up.
+    ++num_frames_;
+    ++curr_num_frame_pairs_;
+
+    // If we've got too many, push up the start of the queue.
+    if (curr_num_frame_pairs_ > kNumFrames) {
+      first_frame_index_ = GetNthIndexFromStart(1);
+      --curr_num_frame_pairs_;
+    }
+  }
+
+  inline int GetNthIndexFromStart(const int offset) const {
+    SCHECK(offset >= 0 && offset < curr_num_frame_pairs_,
+          "Offset out of range!  %d out of %d.", offset, curr_num_frame_pairs_);
+    return (first_frame_index_ + offset) % kNumFrames;
+  }
+
+  void TrackObjects();
+
+  const std::unique_ptr<const TrackerConfig> config_;
+
+  const int frame_width_;
+  const int frame_height_;
+
+  int64_t curr_time_;
+
+  int num_frames_;
+
+  TrackedObjectMap objects_;
+
+  FlowCache flow_cache_;
+
+  KeypointDetector keypoint_detector_;
+
+  int curr_num_frame_pairs_;
+  int first_frame_index_;
+
+  std::unique_ptr<ImageData> frame1_;
+  std::unique_ptr<ImageData> frame2_;
+
+  FramePair frame_pairs_[kNumFrames];
+
+  std::unique_ptr<ObjectDetectorBase> detector_;
+
+  int num_detected_;
+
+ private:
+  void TrackTarget(TrackedObject* const object);
+
+  bool GetBestObjectForDetection(
+      const Detection& detection, TrackedObject** match) const;
+
+  void ProcessDetections(std::vector<Detection>* const detections);
+
+  void DetectTargets();
+
+  // Temp object used in ObjectTracker::CreateNewExample.
+  mutable std::vector<BoundingSquare> squares;
+
+  friend std::ostream& operator<<(std::ostream& stream,
+                                  const ObjectTracker& tracker);
+
+  ObjectTracker(const ObjectTracker&) = delete;
+  void operator=(const ObjectTracker&) = delete;
+};
+
+inline std::ostream& operator<<(std::ostream& stream,
+                                const ObjectTracker& tracker) {
+  stream << "Frame size: " << tracker.frame_width_ << "x"
+         << tracker.frame_height_ << std::endl;
+
+  stream << "Num frames: " << tracker.num_frames_ << std::endl;
+
+  stream << "Curr time: " << tracker.curr_time_ << std::endl;
+
+  const int first_frame_index = tracker.GetNthIndexFromStart(0);
+  const FramePair& first_frame_pair = tracker.frame_pairs_[first_frame_index];
+
+  const int last_frame_index = tracker.GetNthIndexFromEnd(0);
+  const FramePair& last_frame_pair = tracker.frame_pairs_[last_frame_index];
+
+  stream << "first frame: " << first_frame_index << ","
+         << first_frame_pair.end_time_ << "    "
+         << "last frame: " << last_frame_index << ","
+         << last_frame_pair.end_time_ << "   diff: "
+         << last_frame_pair.end_time_ - first_frame_pair.end_time_ << "ms"
+         << std::endl;
+
+  stream << "Tracked targets:";
+  stream << tracker.objects_;
+
+  return stream;
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OBJECT_TRACKER_H_

tensorflow/tools/android/test/jni/object_tracking/object_tracker_jni.cc

@@ -0,0 +1,459 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <android/log.h>
+#include <jni.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <cstdint>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/jni_utils.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/object_tracker.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+
+namespace tf_tracking {
+
+#define OBJECT_TRACKER_METHOD(METHOD_NAME) \
+  Java_org_tensorflow_demo_tracking_ObjectTracker_##METHOD_NAME  // NOLINT
+
+JniLongField object_tracker_field("nativeObjectTracker");
+
+ObjectTracker* get_object_tracker(JNIEnv* env, jobject thiz) {
+  ObjectTracker* const object_tracker =
+      reinterpret_cast<ObjectTracker*>(object_tracker_field.get(env, thiz));
+  CHECK_ALWAYS(object_tracker != NULL, "null object tracker!");
+  return object_tracker;
+}
+
+void set_object_tracker(JNIEnv* env, jobject thiz,
+                        const ObjectTracker* object_tracker) {
+  object_tracker_field.set(env, thiz,
+                           reinterpret_cast<intptr_t>(object_tracker));
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(initNative)(JNIEnv* env, jobject thiz,
+                                               jint width, jint height,
+                                               jboolean always_track);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(releaseMemoryNative)(JNIEnv* env,
+                                                        jobject thiz);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(registerNewObjectWithAppearanceNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloat x1, jfloat y1,
+    jfloat x2, jfloat y2, jbyteArray frame_data);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(setPreviousPositionNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloat x1, jfloat y1,
+    jfloat x2, jfloat y2, jlong timestamp);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(setCurrentPositionNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloat x1, jfloat y1,
+    jfloat x2, jfloat y2);
+
+JNIEXPORT
+jboolean JNICALL OBJECT_TRACKER_METHOD(haveObject)(JNIEnv* env, jobject thiz,
+                                                   jstring object_id);
+
+JNIEXPORT
+jboolean JNICALL OBJECT_TRACKER_METHOD(isObjectVisible)(JNIEnv* env,
+                                                        jobject thiz,
+                                                        jstring object_id);
+
+JNIEXPORT
+jstring JNICALL OBJECT_TRACKER_METHOD(getModelIdNative)(JNIEnv* env,
+                                                        jobject thiz,
+                                                        jstring object_id);
+
+JNIEXPORT
+jfloat JNICALL OBJECT_TRACKER_METHOD(getCurrentCorrelation)(JNIEnv* env,
+                                                            jobject thiz,
+                                                            jstring object_id);
+
+JNIEXPORT
+jfloat JNICALL OBJECT_TRACKER_METHOD(getMatchScore)(JNIEnv* env, jobject thiz,
+                                                    jstring object_id);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(getTrackedPositionNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloatArray rect_array);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(nextFrameNative)(JNIEnv* env, jobject thiz,
+                                                    jbyteArray y_data,
+                                                    jbyteArray uv_data,
+                                                    jlong timestamp,
+                                                    jfloatArray vg_matrix_2x3);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(forgetNative)(JNIEnv* env, jobject thiz,
+                                                 jstring object_id);
+
+JNIEXPORT
+jbyteArray JNICALL OBJECT_TRACKER_METHOD(getKeypointsPacked)(
+    JNIEnv* env, jobject thiz, jfloat scale_factor);
+
+JNIEXPORT
+jfloatArray JNICALL OBJECT_TRACKER_METHOD(getKeypointsNative)(
+    JNIEnv* env, jobject thiz, jboolean only_found_);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(getCurrentPositionNative)(
+    JNIEnv* env, jobject thiz, jlong timestamp, jfloat position_x1,
+    jfloat position_y1, jfloat position_x2, jfloat position_y2,
+    jfloatArray delta);
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(drawNative)(JNIEnv* env, jobject obj,
+                                               jint view_width,
+                                               jint view_height,
+                                               jfloatArray delta);
+
+JNIEXPORT void JNICALL OBJECT_TRACKER_METHOD(downsampleImageNative)(
+    JNIEnv* env, jobject thiz, jint width, jint height, jint row_stride,
+    jbyteArray input, jint factor, jbyteArray output);
+
+#ifdef __cplusplus
+}
+#endif
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(initNative)(JNIEnv* env, jobject thiz,
+                                               jint width, jint height,
+                                               jboolean always_track) {
+  LOGI("Initializing object tracker. %dx%d @%p", width, height, thiz);
+  const Size image_size(width, height);
+  TrackerConfig* const tracker_config = new TrackerConfig(image_size);
+  tracker_config->always_track = always_track;
+
+  // XXX detector
+  ObjectTracker* const tracker = new ObjectTracker(tracker_config, NULL);
+  set_object_tracker(env, thiz, tracker);
+  LOGI("Initialized!");
+
+  CHECK_ALWAYS(get_object_tracker(env, thiz) == tracker,
+               "Failure to set hand tracker!");
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(releaseMemoryNative)(JNIEnv* env,
+                                                        jobject thiz) {
+  delete get_object_tracker(env, thiz);
+  set_object_tracker(env, thiz, NULL);
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(registerNewObjectWithAppearanceNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloat x1, jfloat y1,
+    jfloat x2, jfloat y2, jbyteArray frame_data) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  LOGI("Registering the position of %s at %.2f,%.2f,%.2f,%.2f", id_str, x1, y1,
+       x2, y2);
+
+  jboolean iCopied = JNI_FALSE;
+
+  // Copy image into currFrame.
+  jbyte* pixels = env->GetByteArrayElements(frame_data, &iCopied);
+
+  BoundingBox bounding_box(x1, y1, x2, y2);
+  get_object_tracker(env, thiz)->RegisterNewObjectWithAppearance(
+      id_str, reinterpret_cast<const uint8_t*>(pixels), bounding_box);
+
+  env->ReleaseByteArrayElements(frame_data, pixels, JNI_ABORT);
+
+  env->ReleaseStringUTFChars(object_id, id_str);
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(setPreviousPositionNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloat x1, jfloat y1,
+    jfloat x2, jfloat y2, jlong timestamp) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  LOGI(
+      "Registering the position of %s at %.2f,%.2f,%.2f,%.2f"
+      " at time %lld",
+      id_str, x1, y1, x2, y2, static_cast<long long>(timestamp));
+
+  get_object_tracker(env, thiz)->SetPreviousPositionOfObject(
+      id_str, BoundingBox(x1, y1, x2, y2), timestamp);
+
+  env->ReleaseStringUTFChars(object_id, id_str);
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(setCurrentPositionNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloat x1, jfloat y1,
+    jfloat x2, jfloat y2) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  LOGI("Registering the position of %s at %.2f,%.2f,%.2f,%.2f", id_str, x1, y1,
+       x2, y2);
+
+  get_object_tracker(env, thiz)->SetCurrentPositionOfObject(
+      id_str, BoundingBox(x1, y1, x2, y2));
+
+  env->ReleaseStringUTFChars(object_id, id_str);
+}
+
+JNIEXPORT
+jboolean JNICALL OBJECT_TRACKER_METHOD(haveObject)(JNIEnv* env, jobject thiz,
+                                                   jstring object_id) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  const bool haveObject = get_object_tracker(env, thiz)->HaveObject(id_str);
+  env->ReleaseStringUTFChars(object_id, id_str);
+  return haveObject;
+}
+
+JNIEXPORT
+jboolean JNICALL OBJECT_TRACKER_METHOD(isObjectVisible)(JNIEnv* env,
+                                                        jobject thiz,
+                                                        jstring object_id) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  const bool visible = get_object_tracker(env, thiz)->IsObjectVisible(id_str);
+  env->ReleaseStringUTFChars(object_id, id_str);
+  return visible;
+}
+
+JNIEXPORT
+jstring JNICALL OBJECT_TRACKER_METHOD(getModelIdNative)(JNIEnv* env,
+                                                        jobject thiz,
+                                                        jstring object_id) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+  const TrackedObject* const object =
+      get_object_tracker(env, thiz)->GetObject(id_str);
+  env->ReleaseStringUTFChars(object_id, id_str);
+  jstring model_name = env->NewStringUTF(object->GetModel()->GetName().c_str());
+  return model_name;
+}
+
+JNIEXPORT
+jfloat JNICALL OBJECT_TRACKER_METHOD(getCurrentCorrelation)(JNIEnv* env,
+                                                            jobject thiz,
+                                                            jstring object_id) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  const float correlation =
+      get_object_tracker(env, thiz)->GetObject(id_str)->GetCorrelation();
+  env->ReleaseStringUTFChars(object_id, id_str);
+  return correlation;
+}
+
+JNIEXPORT
+jfloat JNICALL OBJECT_TRACKER_METHOD(getMatchScore)(JNIEnv* env, jobject thiz,
+                                                    jstring object_id) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  const float match_score =
+      get_object_tracker(env, thiz)->GetObject(id_str)->GetMatchScore().value;
+  env->ReleaseStringUTFChars(object_id, id_str);
+  return match_score;
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(getTrackedPositionNative)(
+    JNIEnv* env, jobject thiz, jstring object_id, jfloatArray rect_array) {
+  jboolean iCopied = JNI_FALSE;
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  const BoundingBox bounding_box =
+      get_object_tracker(env, thiz)->GetObject(id_str)->GetPosition();
+  env->ReleaseStringUTFChars(object_id, id_str);
+
+  jfloat* rect = env->GetFloatArrayElements(rect_array, &iCopied);
+  bounding_box.CopyToArray(reinterpret_cast<float*>(rect));
+  env->ReleaseFloatArrayElements(rect_array, rect, 0);
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(nextFrameNative)(JNIEnv* env, jobject thiz,
+                                                    jbyteArray y_data,
+                                                    jbyteArray uv_data,
+                                                    jlong timestamp,
+                                                    jfloatArray vg_matrix_2x3) {
+  TimeLog("Starting object tracker");
+
+  jboolean iCopied = JNI_FALSE;
+
+  float vision_gyro_matrix_array[6];
+  jfloat* jmat = NULL;
+
+  if (vg_matrix_2x3 != NULL) {
+    // Copy the alignment matrix into a float array.
+    jmat = env->GetFloatArrayElements(vg_matrix_2x3, &iCopied);
+    for (int i = 0; i < 6; ++i) {
+      vision_gyro_matrix_array[i] = static_cast<float>(jmat[i]);
+    }
+  }
+  // Copy image into currFrame.
+  jbyte* pixels = env->GetByteArrayElements(y_data, &iCopied);
+  jbyte* uv_pixels =
+      uv_data != NULL ? env->GetByteArrayElements(uv_data, &iCopied) : NULL;
+
+  TimeLog("Got elements");
+
+  // Add the frame to the object tracker object.
+  get_object_tracker(env, thiz)->NextFrame(
+      reinterpret_cast<uint8_t*>(pixels), reinterpret_cast<uint8_t*>(uv_pixels),
+      timestamp, vg_matrix_2x3 != NULL ? vision_gyro_matrix_array : NULL);
+
+  env->ReleaseByteArrayElements(y_data, pixels, JNI_ABORT);
+
+  if (uv_data != NULL) {
+    env->ReleaseByteArrayElements(uv_data, uv_pixels, JNI_ABORT);
+  }
+
+  if (vg_matrix_2x3 != NULL) {
+    env->ReleaseFloatArrayElements(vg_matrix_2x3, jmat, JNI_ABORT);
+  }
+
+  TimeLog("Released elements");
+
+  PrintTimeLog();
+  ResetTimeLog();
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(forgetNative)(JNIEnv* env, jobject thiz,
+                                                 jstring object_id) {
+  const char* const id_str = env->GetStringUTFChars(object_id, 0);
+
+  get_object_tracker(env, thiz)->ForgetTarget(id_str);
+
+  env->ReleaseStringUTFChars(object_id, id_str);
+}
+
+JNIEXPORT
+jfloatArray JNICALL OBJECT_TRACKER_METHOD(getKeypointsNative)(
+    JNIEnv* env, jobject thiz, jboolean only_found) {
+  jfloat keypoint_arr[kMaxKeypoints * kKeypointStep];
+
+  const int number_of_keypoints =
+      get_object_tracker(env, thiz)->GetKeypoints(only_found, keypoint_arr);
+
+  // Create and return the array that will be passed back to Java.
+  jfloatArray keypoints =
+      env->NewFloatArray(number_of_keypoints * kKeypointStep);
+  if (keypoints == NULL) {
+    LOGE("null array!");
+    return NULL;
+  }
+  env->SetFloatArrayRegion(keypoints, 0, number_of_keypoints * kKeypointStep,
+                           keypoint_arr);
+
+  return keypoints;
+}
+
+JNIEXPORT
+jbyteArray JNICALL OBJECT_TRACKER_METHOD(getKeypointsPacked)(
+    JNIEnv* env, jobject thiz, jfloat scale_factor) {
+  // 2 bytes to a uint16_t and two pairs of xy coordinates per keypoint.
+  const int bytes_per_keypoint = sizeof(uint16_t) * 2 * 2;
+  jbyte keypoint_arr[kMaxKeypoints * bytes_per_keypoint];
+
+  const int number_of_keypoints =
+      get_object_tracker(env, thiz)->GetKeypointsPacked(
+          reinterpret_cast<uint16_t*>(keypoint_arr), scale_factor);
+
+  // Create and return the array that will be passed back to Java.
+  jbyteArray keypoints =
+      env->NewByteArray(number_of_keypoints * bytes_per_keypoint);
+
+  if (keypoints == NULL) {
+    LOGE("null array!");
+    return NULL;
+  }
+
+  env->SetByteArrayRegion(
+      keypoints, 0, number_of_keypoints * bytes_per_keypoint, keypoint_arr);
+
+  return keypoints;
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(getCurrentPositionNative)(
+    JNIEnv* env, jobject thiz, jlong timestamp, jfloat position_x1,
+    jfloat position_y1, jfloat position_x2, jfloat position_y2,
+    jfloatArray delta) {
+  jfloat point_arr[4];
+
+  const BoundingBox new_position = get_object_tracker(env, thiz)->TrackBox(
+      BoundingBox(position_x1, position_y1, position_x2, position_y2),
+      timestamp);
+
+  new_position.CopyToArray(point_arr);
+  env->SetFloatArrayRegion(delta, 0, 4, point_arr);
+}
+
+JNIEXPORT
+void JNICALL OBJECT_TRACKER_METHOD(drawNative)(
+    JNIEnv* env, jobject thiz, jint view_width, jint view_height,
+    jfloatArray frame_to_canvas_arr) {
+  ObjectTracker* object_tracker = get_object_tracker(env, thiz);
+  if (object_tracker != NULL) {
+    jfloat* frame_to_canvas =
+        env->GetFloatArrayElements(frame_to_canvas_arr, NULL);
+
+    object_tracker->Draw(view_width, view_height, frame_to_canvas);
+    env->ReleaseFloatArrayElements(frame_to_canvas_arr, frame_to_canvas,
+                                   JNI_ABORT);
+  }
+}
+
+JNIEXPORT void JNICALL OBJECT_TRACKER_METHOD(downsampleImageNative)(
+    JNIEnv* env, jobject thiz, jint width, jint height, jint row_stride,
+    jbyteArray input, jint factor, jbyteArray output) {
+  if (input == NULL || output == NULL) {
+    LOGW("Received null arrays, hopefully this is a test!");
+    return;
+  }
+
+  jbyte* const input_array = env->GetByteArrayElements(input, 0);
+  jbyte* const output_array = env->GetByteArrayElements(output, 0);
+
+  {
+    tf_tracking::Image<uint8_t> full_image(
+        width, height, reinterpret_cast<uint8_t*>(input_array), false);
+
+    const int new_width = (width + factor - 1) / factor;
+    const int new_height = (height + factor - 1) / factor;
+
+    tf_tracking::Image<uint8_t> downsampled_image(
+        new_width, new_height, reinterpret_cast<uint8_t*>(output_array), false);
+
+    downsampled_image.DownsampleAveraged(
+        reinterpret_cast<uint8_t*>(input_array), row_stride, factor);
+  }
+
+  env->ReleaseByteArrayElements(input, input_array, JNI_ABORT);
+  env->ReleaseByteArrayElements(output, output_array, 0);
+}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/optical_flow.cc

@@ -0,0 +1,482 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/tools/android/test/jni/object_tracking/optical_flow.h"
+
+#include <math.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/flow_cache.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/frame_pair.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_data.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint_detector.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+OpticalFlow::OpticalFlow(const OpticalFlowConfig* const config)
+    : config_(config),
+      frame1_(NULL),
+      frame2_(NULL),
+      working_size_(config->image_size) {}
+
+
+void OpticalFlow::NextFrame(const ImageData* const image_data) {
+  // Special case for the first frame: make sure the image ends up in
+  // frame1_ so that keypoint detection can be done on it if desired.
+  frame1_ = (frame1_ == NULL) ? image_data : frame2_;
+  frame2_ = image_data;
+}
+
+
+// Static heart of the optical flow computation.
+// Lucas Kanade algorithm.
+bool OpticalFlow::FindFlowAtPoint_LK(const Image<uint8_t>& img_I,
+                                     const Image<uint8_t>& img_J,
+                                     const Image<int32_t>& I_x,
+                                     const Image<int32_t>& I_y, const float p_x,
+                                     const float p_y, float* out_g_x,
+                                     float* out_g_y) {
+  float g_x = *out_g_x;
+  float g_y = *out_g_y;
+  // Get values for frame 1.  They remain constant through the inner
+  // iteration loop.
+  float vals_I[kFlowArraySize];
+  float vals_I_x[kFlowArraySize];
+  float vals_I_y[kFlowArraySize];
+
+  const int kPatchSize = 2 * kFlowIntegrationWindowSize + 1;
+  const float kWindowSizeFloat = static_cast<float>(kFlowIntegrationWindowSize);
+
+#if USE_FIXED_POINT_FLOW
+  const int fixed_x_max = RealToFixed1616(img_I.width_less_one_) - 1;
+  const int fixed_y_max = RealToFixed1616(img_I.height_less_one_) - 1;
+#else
+  const float real_x_max = I_x.width_less_one_ - EPSILON;
+  const float real_y_max = I_x.height_less_one_ - EPSILON;
+#endif
+
+  // Get the window around the original point.
+  const float src_left_real = p_x - kWindowSizeFloat;
+  const float src_top_real = p_y - kWindowSizeFloat;
+  float* vals_I_ptr = vals_I;
+  float* vals_I_x_ptr = vals_I_x;
+  float* vals_I_y_ptr = vals_I_y;
+#if USE_FIXED_POINT_FLOW
+  // Source integer coordinates.
+  const int src_left_fixed = RealToFixed1616(src_left_real);
+  const int src_top_fixed = RealToFixed1616(src_top_real);
+
+  for (int y = 0; y < kPatchSize; ++y) {
+    const int fp_y = Clip(src_top_fixed + (y << 16), 0, fixed_y_max);
+
+    for (int x = 0; x < kPatchSize; ++x) {
+      const int fp_x = Clip(src_left_fixed + (x << 16), 0, fixed_x_max);
+
+      *vals_I_ptr++ = img_I.GetPixelInterpFixed1616(fp_x, fp_y);
+      *vals_I_x_ptr++ = I_x.GetPixelInterpFixed1616(fp_x, fp_y);
+      *vals_I_y_ptr++ = I_y.GetPixelInterpFixed1616(fp_x, fp_y);
+    }
+  }
+#else
+  for (int y = 0; y < kPatchSize; ++y) {
+    const float y_pos = Clip(src_top_real + y, 0.0f, real_y_max);
+
+    for (int x = 0; x < kPatchSize; ++x) {
+      const float x_pos = Clip(src_left_real + x, 0.0f, real_x_max);
+
+      *vals_I_ptr++ = img_I.GetPixelInterp(x_pos, y_pos);
+      *vals_I_x_ptr++ = I_x.GetPixelInterp(x_pos, y_pos);
+      *vals_I_y_ptr++ = I_y.GetPixelInterp(x_pos, y_pos);
+    }
+  }
+#endif
+
+  // Compute the spatial gradient matrix about point p.
+  float G[] = { 0, 0, 0, 0 };
+  CalculateG(vals_I_x, vals_I_y, kFlowArraySize, G);
+
+  // Find the inverse of G.
+  float G_inv[4];
+  if (!Invert2x2(G, G_inv)) {
+    return false;
+  }
+
+#if NORMALIZE
+  const float mean_I = ComputeMean(vals_I, kFlowArraySize);
+  const float std_dev_I = ComputeStdDev(vals_I, kFlowArraySize, mean_I);
+#endif
+
+  // Iterate kNumIterations times or until we converge.
+  for (int iteration = 0; iteration < kNumIterations; ++iteration) {
+    // Get values for frame 2.
+    float vals_J[kFlowArraySize];
+
+    // Get the window around the destination point.
+    const float left_real = p_x + g_x - kWindowSizeFloat;
+    const float top_real  = p_y + g_y - kWindowSizeFloat;
+    float* vals_J_ptr = vals_J;
+#if USE_FIXED_POINT_FLOW
+    // The top-left sub-pixel is set for the current iteration (in 16:16
+    // fixed). This is constant over one iteration.
+    const int left_fixed = RealToFixed1616(left_real);
+    const int top_fixed  = RealToFixed1616(top_real);
+
+    for (int win_y = 0; win_y < kPatchSize; ++win_y) {
+      const int fp_y = Clip(top_fixed + (win_y << 16), 0, fixed_y_max);
+      for (int win_x = 0; win_x < kPatchSize; ++win_x) {
+        const int fp_x = Clip(left_fixed + (win_x << 16), 0, fixed_x_max);
+        *vals_J_ptr++ = img_J.GetPixelInterpFixed1616(fp_x, fp_y);
+      }
+    }
+#else
+    for (int win_y = 0; win_y < kPatchSize; ++win_y) {
+      const float y_pos = Clip(top_real + win_y, 0.0f, real_y_max);
+      for (int win_x = 0; win_x < kPatchSize; ++win_x) {
+        const float x_pos = Clip(left_real + win_x, 0.0f, real_x_max);
+        *vals_J_ptr++ = img_J.GetPixelInterp(x_pos, y_pos);
+      }
+    }
+#endif
+
+#if NORMALIZE
+    const float mean_J = ComputeMean(vals_J, kFlowArraySize);
+    const float std_dev_J = ComputeStdDev(vals_J, kFlowArraySize, mean_J);
+
+    // TODO(andrewharp): Probably better to completely detect and handle the
+    // "corner case" where the patch is fully outside the image diagonally.
+    const float std_dev_ratio = std_dev_J > 0.0f ? std_dev_I / std_dev_J : 1.0f;
+#endif
+
+    // Compute image mismatch vector.
+    float b_x = 0.0f;
+    float b_y = 0.0f;
+
+    vals_I_ptr = vals_I;
+    vals_J_ptr = vals_J;
+    vals_I_x_ptr = vals_I_x;
+    vals_I_y_ptr = vals_I_y;
+
+    for (int win_y = 0; win_y < kPatchSize; ++win_y) {
+      for (int win_x = 0; win_x < kPatchSize; ++win_x) {
+#if NORMALIZE
+        // Normalized Image difference.
+        const float dI =
+            (*vals_I_ptr++ - mean_I) - (*vals_J_ptr++ - mean_J) * std_dev_ratio;
+#else
+        const float dI = *vals_I_ptr++ - *vals_J_ptr++;
+#endif
+        b_x += dI * *vals_I_x_ptr++;
+        b_y += dI * *vals_I_y_ptr++;
+      }
+    }
+
+    // Optical flow... solve n = G^-1 * b
+    const float n_x = (G_inv[0] * b_x) + (G_inv[1] * b_y);
+    const float n_y = (G_inv[2] * b_x) + (G_inv[3] * b_y);
+
+    // Update best guess with residual displacement from this level and
+    // iteration.
+    g_x += n_x;
+    g_y += n_y;
+
+    // LOGV("Iteration %d: delta (%.3f, %.3f)", iteration, n_x, n_y);
+
+    // Abort early if we're already below the threshold.
+    if (Square(n_x) + Square(n_y) < Square(kTrackingAbortThreshold)) {
+      break;
+    }
+  }  // Iteration.
+
+  // Copy value back into output.
+  *out_g_x = g_x;
+  *out_g_y = g_y;
+  return true;
+}
+
+
+// Pointwise flow using translational 2dof ESM.
+bool OpticalFlow::FindFlowAtPoint_ESM(
+    const Image<uint8_t>& img_I, const Image<uint8_t>& img_J,
+    const Image<int32_t>& I_x, const Image<int32_t>& I_y,
+    const Image<int32_t>& J_x, const Image<int32_t>& J_y, const float p_x,
+    const float p_y, float* out_g_x, float* out_g_y) {
+  float g_x = *out_g_x;
+  float g_y = *out_g_y;
+  const float area_inv = 1.0f / static_cast<float>(kFlowArraySize);
+
+  // Get values for frame 1. They remain constant through the inner
+  // iteration loop.
+  uint8_t vals_I[kFlowArraySize];
+  uint8_t vals_J[kFlowArraySize];
+  int16_t src_gradient_x[kFlowArraySize];
+  int16_t src_gradient_y[kFlowArraySize];
+
+  // TODO(rspring): try out the IntegerPatchAlign() method once
+  // the code for that is in ../common.
+  const float wsize_float = static_cast<float>(kFlowIntegrationWindowSize);
+  const int src_left_fixed = RealToFixed1616(p_x - wsize_float);
+  const int src_top_fixed = RealToFixed1616(p_y - wsize_float);
+  const int patch_size = 2 * kFlowIntegrationWindowSize + 1;
+
+  // Create the keypoint template patch from a subpixel location.
+  if (!img_I.ExtractPatchAtSubpixelFixed1616(src_left_fixed, src_top_fixed,
+                                             patch_size, patch_size, vals_I) ||
+      !I_x.ExtractPatchAtSubpixelFixed1616(src_left_fixed, src_top_fixed,
+                                           patch_size, patch_size,
+                                           src_gradient_x) ||
+      !I_y.ExtractPatchAtSubpixelFixed1616(src_left_fixed, src_top_fixed,
+                                           patch_size, patch_size,
+                                           src_gradient_y)) {
+    return false;
+  }
+
+  int bright_offset = 0;
+  int sum_diff = 0;
+
+  // The top-left sub-pixel is set for the current iteration (in 16:16 fixed).
+  // This is constant over one iteration.
+  int left_fixed = RealToFixed1616(p_x + g_x - wsize_float);
+  int top_fixed  = RealToFixed1616(p_y + g_y - wsize_float);
+
+  // The truncated version gives the most top-left pixel that is used.
+  int left_trunc = left_fixed >> 16;
+  int top_trunc = top_fixed >> 16;
+
+  // Compute an initial brightness offset.
+  if (kDoBrightnessNormalize &&
+      left_trunc >= 0 && top_trunc >= 0 &&
+      (left_trunc + patch_size) < img_J.width_less_one_ &&
+      (top_trunc + patch_size) < img_J.height_less_one_) {
+    int templ_index = 0;
+    const uint8_t* j_row = img_J[top_trunc] + left_trunc;
+
+    const int j_stride = img_J.stride();
+
+    for (int y = 0; y < patch_size; ++y, j_row += j_stride) {
+      for (int x = 0; x < patch_size; ++x) {
+        sum_diff += static_cast<int>(j_row[x]) - vals_I[templ_index++];
+      }
+    }
+
+    bright_offset = static_cast<int>(static_cast<float>(sum_diff) * area_inv);
+  }
+
+  // Iterate kNumIterations times or until we go out of image.
+  for (int iteration = 0; iteration < kNumIterations; ++iteration) {
+    int jtj[3] = { 0, 0, 0 };
+    int jtr[2] = { 0, 0 };
+    sum_diff = 0;
+
+    // Extract the target image values.
+    // Extract the gradient from the target image patch and accumulate to
+    // the gradient of the source image patch.
+    if (!img_J.ExtractPatchAtSubpixelFixed1616(left_fixed, top_fixed,
+                                               patch_size, patch_size,
+                                               vals_J)) {
+      break;
+    }
+
+    const uint8_t* templ_row = vals_I;
+    const uint8_t* extract_row = vals_J;
+    const int16_t* src_dx_row = src_gradient_x;
+    const int16_t* src_dy_row = src_gradient_y;
+
+    for (int y = 0; y < patch_size; ++y, templ_row += patch_size,
+         src_dx_row += patch_size, src_dy_row += patch_size,
+         extract_row += patch_size) {
+      const int fp_y = top_fixed + (y << 16);
+      for (int x = 0; x < patch_size; ++x) {
+        const int fp_x = left_fixed + (x << 16);
+        int32_t target_dx = J_x.GetPixelInterpFixed1616(fp_x, fp_y);
+        int32_t target_dy = J_y.GetPixelInterpFixed1616(fp_x, fp_y);
+
+        // Combine the two Jacobians.
+        // Right-shift by one to account for the fact that we add
+        // two Jacobians.
+        int32_t dx = (src_dx_row[x] + target_dx) >> 1;
+        int32_t dy = (src_dy_row[x] + target_dy) >> 1;
+
+        // The current residual b - h(q) == extracted - (template + offset)
+        int32_t diff = static_cast<int32_t>(extract_row[x]) -
+                       static_cast<int32_t>(templ_row[x]) - bright_offset;
+
+        jtj[0] += dx * dx;
+        jtj[1] += dx * dy;
+        jtj[2] += dy * dy;
+
+        jtr[0] += dx * diff;
+        jtr[1] += dy * diff;
+
+        sum_diff += diff;
+      }
+    }
+
+    const float jtr1_float = static_cast<float>(jtr[0]);
+    const float jtr2_float = static_cast<float>(jtr[1]);
+
+    // Add some baseline stability to the system.
+    jtj[0] += kEsmRegularizer;
+    jtj[2] += kEsmRegularizer;
+
+    const int64_t prod1 = static_cast<int64_t>(jtj[0]) * jtj[2];
+    const int64_t prod2 = static_cast<int64_t>(jtj[1]) * jtj[1];
+
+    // One ESM step.
+    const float jtj_1[4] = { static_cast<float>(jtj[2]),
+                             static_cast<float>(-jtj[1]),
+                             static_cast<float>(-jtj[1]),
+                             static_cast<float>(jtj[0]) };
+    const double det_inv = 1.0 / static_cast<double>(prod1 - prod2);
+
+    g_x -= det_inv * (jtj_1[0] * jtr1_float + jtj_1[1] * jtr2_float);
+    g_y -= det_inv * (jtj_1[2] * jtr1_float + jtj_1[3] * jtr2_float);
+
+    if (kDoBrightnessNormalize) {
+      bright_offset +=
+          static_cast<int>(area_inv * static_cast<float>(sum_diff) + 0.5f);
+    }
+
+    // Update top left position.
+    left_fixed = RealToFixed1616(p_x + g_x - wsize_float);
+    top_fixed  = RealToFixed1616(p_y + g_y - wsize_float);
+
+    left_trunc = left_fixed >> 16;
+    top_trunc = top_fixed >> 16;
+
+    // Abort iterations if we go out of borders.
+    if (left_trunc < 0 || top_trunc < 0 ||
+        (left_trunc + patch_size) >= J_x.width_less_one_ ||
+        (top_trunc + patch_size) >= J_y.height_less_one_) {
+      break;
+    }
+  }  // Iteration.
+
+  // Copy value back into output.
+  *out_g_x = g_x;
+  *out_g_y = g_y;
+  return true;
+}
+
+
+bool OpticalFlow::FindFlowAtPointReversible(
+    const int level, const float u_x, const float u_y,
+    const bool reverse_flow,
+    float* flow_x, float* flow_y) const {
+  const ImageData& frame_a = reverse_flow ? *frame2_ : *frame1_;
+  const ImageData& frame_b = reverse_flow ? *frame1_ : *frame2_;
+
+  // Images I (prev) and J (next).
+  const Image<uint8_t>& img_I = *frame_a.GetPyramidSqrt2Level(level * 2);
+  const Image<uint8_t>& img_J = *frame_b.GetPyramidSqrt2Level(level * 2);
+
+  // Computed gradients.
+  const Image<int32_t>& I_x = *frame_a.GetSpatialX(level);
+  const Image<int32_t>& I_y = *frame_a.GetSpatialY(level);
+  const Image<int32_t>& J_x = *frame_b.GetSpatialX(level);
+  const Image<int32_t>& J_y = *frame_b.GetSpatialY(level);
+
+  // Shrink factor from original.
+  const float shrink_factor = (1 << level);
+
+  // Image position vector (p := u^l), scaled for this level.
+  const float scaled_p_x = u_x / shrink_factor;
+  const float scaled_p_y = u_y / shrink_factor;
+
+  float scaled_flow_x = *flow_x / shrink_factor;
+  float scaled_flow_y = *flow_y / shrink_factor;
+
+  // LOGE("FindFlowAtPoint level %d: %5.2f, %5.2f (%5.2f, %5.2f)", level,
+  //     scaled_p_x, scaled_p_y, &scaled_flow_x, &scaled_flow_y);
+
+  const bool success = kUseEsm ?
+    FindFlowAtPoint_ESM(img_I, img_J, I_x, I_y, J_x, J_y,
+                        scaled_p_x, scaled_p_y,
+                        &scaled_flow_x, &scaled_flow_y) :
+    FindFlowAtPoint_LK(img_I, img_J, I_x, I_y,
+                       scaled_p_x, scaled_p_y,
+                       &scaled_flow_x, &scaled_flow_y);
+
+  *flow_x = scaled_flow_x * shrink_factor;
+  *flow_y = scaled_flow_y * shrink_factor;
+
+  return success;
+}
+
+
+bool OpticalFlow::FindFlowAtPointSingleLevel(
+    const int level,
+    const float u_x, const float u_y,
+    const bool filter_by_fb_error,
+    float* flow_x, float* flow_y) const {
+  if (!FindFlowAtPointReversible(level, u_x, u_y, false, flow_x, flow_y)) {
+    return false;
+  }
+
+  if (filter_by_fb_error) {
+    const float new_position_x = u_x + *flow_x;
+    const float new_position_y = u_y + *flow_y;
+
+    float reverse_flow_x = 0.0f;
+    float reverse_flow_y = 0.0f;
+
+    // Now find the backwards flow and confirm it lines up with the original
+    // starting point.
+    if (!FindFlowAtPointReversible(level, new_position_x, new_position_y,
+                                   true,
+                                   &reverse_flow_x, &reverse_flow_y)) {
+      LOGE("Backward error!");
+      return false;
+    }
+
+    const float discrepancy_length =
+        sqrtf(Square(*flow_x + reverse_flow_x) +
+              Square(*flow_y + reverse_flow_y));
+
+    const float flow_length = sqrtf(Square(*flow_x) + Square(*flow_y));
+
+    return discrepancy_length <
+        (kMaxForwardBackwardErrorAllowed * flow_length);
+  }
+
+  return true;
+}
+
+
+// An implementation of the Pyramidal Lucas-Kanade Optical Flow algorithm.
+// See http://robots.stanford.edu/cs223b04/algo_tracking.pdf for details.
+bool OpticalFlow::FindFlowAtPointPyramidal(const float u_x, const float u_y,
+                                           const bool filter_by_fb_error,
+                                           float* flow_x, float* flow_y) const {
+  const int max_level = MAX(kMinNumPyramidLevelsToUseForAdjustment,
+                            kNumPyramidLevels - kNumCacheLevels);
+
+  // For every level in the pyramid, update the coordinates of the best match.
+  for (int l = max_level - 1; l >= 0; --l) {
+    if (!FindFlowAtPointSingleLevel(l, u_x, u_y,
+                                    filter_by_fb_error, flow_x, flow_y)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/optical_flow.h

@@ -0,0 +1,100 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OPTICAL_FLOW_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OPTICAL_FLOW_H_
+
+#include "tensorflow/tools/android/test/jni/object_tracking/config.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/frame_pair.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image_data.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/keypoint.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+class FlowCache;
+
+// Class encapsulating all the data and logic necessary for performing optical
+// flow.
+class OpticalFlow {
+ public:
+  explicit OpticalFlow(const OpticalFlowConfig* const config);
+
+  // Add a new frame to the optical flow.  Will update all the non-keypoint
+  // related member variables.
+  //
+  // new_frame should be a buffer of grayscale values, one byte per pixel,
+  // at the original frame_width and frame_height used to initialize the
+  // OpticalFlow object.  Downsampling will be handled internally.
+  //
+  // time_stamp should be a time in milliseconds that later calls to this and
+  // other methods will be relative to.
+  void NextFrame(const ImageData* const image_data);
+
+  // An implementation of the Lucas-Kanade Optical Flow algorithm.
+  static bool FindFlowAtPoint_LK(const Image<uint8_t>& img_I,
+                                 const Image<uint8_t>& img_J,
+                                 const Image<int32_t>& I_x,
+                                 const Image<int32_t>& I_y, const float p_x,
+                                 const float p_y, float* out_g_x,
+                                 float* out_g_y);
+
+  // Pointwise flow using translational 2dof ESM.
+  static bool FindFlowAtPoint_ESM(
+      const Image<uint8_t>& img_I, const Image<uint8_t>& img_J,
+      const Image<int32_t>& I_x, const Image<int32_t>& I_y,
+      const Image<int32_t>& J_x, const Image<int32_t>& J_y, const float p_x,
+      const float p_y, float* out_g_x, float* out_g_y);
+
+  // Finds the flow using a specific level, in either direction.
+  // If reversed, the coordinates are in the context of the latest
+  // frame, not the frame before it.
+  // All coordinates used in parameters are global, not scaled.
+  bool FindFlowAtPointReversible(
+      const int level, const float u_x, const float u_y,
+      const bool reverse_flow,
+      float* final_x, float* final_y) const;
+
+  // Finds the flow using a specific level, filterable by forward-backward
+  // error. All coordinates used in parameters are global, not scaled.
+  bool FindFlowAtPointSingleLevel(const int level,
+                                  const float u_x, const float u_y,
+                                  const bool filter_by_fb_error,
+                                  float* flow_x, float* flow_y) const;
+
+  // Pyramidal optical-flow using all levels.
+  bool FindFlowAtPointPyramidal(const float u_x, const float u_y,
+                                const bool filter_by_fb_error,
+                                float* flow_x, float* flow_y) const;
+
+ private:
+  const OpticalFlowConfig* const config_;
+
+  const ImageData* frame1_;
+  const ImageData* frame2_;
+
+  // Size of the internally allocated images (after original is downsampled).
+  const Size working_size_;
+
+  OpticalFlow(const OpticalFlow&) = delete;
+  void operator=(const OpticalFlow&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_OPTICAL_FLOW_H_

tensorflow/tools/android/test/jni/object_tracking/sprite.h

@@ -0,0 +1,203 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_SPRITE_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_SPRITE_H_
+
+#ifdef __RENDER_OPENGL__
+
+#include <GLES/gl.h>
+#include <GLES/glext.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+
+namespace tf_tracking {
+
+// This class encapsulates the logic necessary to load an render image data
+// at the same aspect ratio as the original source.
+class Sprite {
+ public:
+  // Only create Sprites when you have an OpenGl context.
+  explicit Sprite(const Image<uint8_t>& image) { LoadTexture(image, NULL); }
+
+  Sprite(const Image<uint8_t>& image, const BoundingBox* const area) {
+    LoadTexture(image, area);
+  }
+
+  // Also, try to only delete a Sprite when holding an OpenGl context.
+  ~Sprite() {
+    glDeleteTextures(1, &texture_);
+  }
+
+  inline int GetWidth() const {
+    return actual_width_;
+  }
+
+  inline int GetHeight() const {
+    return actual_height_;
+  }
+
+  // Draw the sprite at 0,0 - original width/height in the current reference
+  // frame. Any transformations desired must be applied before calling this
+  // function.
+  void Draw() const {
+    const float float_width = static_cast<float>(actual_width_);
+    const float float_height = static_cast<float>(actual_height_);
+
+    // Where it gets rendered to.
+    const float vertices[] = { 0.0f, 0.0f, 0.0f,
+                               0.0f, float_height, 0.0f,
+                               float_width, 0.0f, 0.0f,
+                               float_width, float_height, 0.0f,
+                               };
+
+    // The coordinates the texture gets drawn from.
+    const float max_x = float_width / texture_width_;
+    const float max_y = float_height / texture_height_;
+    const float textureVertices[] = {
+        0, 0,
+        0, max_y,
+        max_x, 0,
+        max_x, max_y,
+    };
+
+    glEnable(GL_TEXTURE_2D);
+    glBindTexture(GL_TEXTURE_2D, texture_);
+
+    glEnableClientState(GL_VERTEX_ARRAY);
+    glEnableClientState(GL_TEXTURE_COORD_ARRAY);
+
+    glVertexPointer(3, GL_FLOAT, 0, vertices);
+    glTexCoordPointer(2, GL_FLOAT, 0, textureVertices);
+
+    glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+    glDisableClientState(GL_VERTEX_ARRAY);
+    glDisableClientState(GL_TEXTURE_COORD_ARRAY);
+  }
+
+ private:
+  inline int GetNextPowerOfTwo(const int number) const {
+    int power_of_two = 1;
+    while (power_of_two < number) {
+      power_of_two *= 2;
+    }
+    return power_of_two;
+  }
+
+  // TODO(andrewharp): Allow sprites to have their textures reloaded.
+  void LoadTexture(const Image<uint8_t>& texture_source,
+                   const BoundingBox* const area) {
+    glEnable(GL_TEXTURE_2D);
+
+    glGenTextures(1, &texture_);
+
+    glBindTexture(GL_TEXTURE_2D, texture_);
+
+    int left = 0;
+    int top = 0;
+
+    if (area != NULL) {
+      // If a sub-region was provided to pull the texture from, use that.
+      left = area->left_;
+      top = area->top_;
+      actual_width_ = area->GetWidth();
+      actual_height_ = area->GetHeight();
+    } else {
+      actual_width_ = texture_source.GetWidth();
+      actual_height_ = texture_source.GetHeight();
+    }
+
+    // The textures must be a power of two, so find the sizes that are large
+    // enough to contain the image data.
+    texture_width_ = GetNextPowerOfTwo(actual_width_);
+    texture_height_ = GetNextPowerOfTwo(actual_height_);
+
+    bool allocated_data = false;
+    uint8_t* texture_data;
+
+    // Except in the lucky case where we're not using a sub-region of the
+    // original image AND the source data has dimensions that are power of two,
+    // care must be taken to copy data at the appropriate source and destination
+    // strides so that the final block can be copied directly into texture
+    // memory.
+    // TODO(andrewharp): Figure out if data can be pulled directly from the
+    // source image with some alignment modifications.
+    if (left != 0 || top != 0 ||
+        actual_width_ != texture_source.GetWidth() ||
+        actual_height_ != texture_source.GetHeight()) {
+      texture_data = new uint8_t[actual_width_ * actual_height_];
+
+      for (int y = 0; y < actual_height_; ++y) {
+        memcpy(texture_data + actual_width_ * y, texture_source[top + y] + left,
+               actual_width_ * sizeof(uint8_t));
+      }
+      allocated_data = true;
+    } else {
+      // Cast away const-ness because for some reason glTexSubImage2D wants
+      // a non-const data pointer.
+      texture_data = const_cast<uint8_t*>(texture_source.data());
+    }
+
+    glTexImage2D(GL_TEXTURE_2D,
+                 0,
+                 GL_LUMINANCE,
+                 texture_width_,
+                 texture_height_,
+                 0,
+                 GL_LUMINANCE,
+                 GL_UNSIGNED_BYTE,
+                 NULL);
+
+    glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
+    glTexSubImage2D(GL_TEXTURE_2D,
+                    0,
+                    0,
+                    0,
+                    actual_width_,
+                    actual_height_,
+                    GL_LUMINANCE,
+                    GL_UNSIGNED_BYTE,
+                    texture_data);
+
+    if (allocated_data) {
+      delete(texture_data);
+    }
+
+    glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+  }
+
+  // The id for the texture on the GPU.
+  GLuint texture_;
+
+  // The width and height to be used for display purposes, referring to the
+  // dimensions of the original texture.
+  int actual_width_;
+  int actual_height_;
+
+  // The allocated dimensions of the texture data, which must be powers of 2.
+  int texture_width_;
+  int texture_height_;
+
+  Sprite(const Sprite&) = delete;
+  void operator=(const Sprite&) = delete;
+};
+
+}  // namespace tf_tracking
+
+#endif  // __RENDER_OPENGL__
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_SPRITE_H_

tensorflow/tools/android/test/jni/object_tracking/time_log.cc

@@ -0,0 +1,25 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/tools/android/test/jni/object_tracking/time_log.h"
+
+#ifdef LOG_TIME
+// Storage for logging functionality.
+int num_time_logs = 0;
+LogEntry time_logs[NUM_LOGS];
+
+int num_avg_entries = 0;
+AverageEntry avg_entries[NUM_LOGS];
+#endif

tensorflow/tools/android/test/jni/object_tracking/time_log.h

@@ -0,0 +1,137 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// Utility functions for performance profiling.
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_TIME_LOG_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_TIME_LOG_H_
+
+#include <stdint.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+#ifdef LOG_TIME
+
+// Blend constant for running average.
+#define ALPHA 0.98f
+#define NUM_LOGS 100
+
+struct LogEntry {
+  const char* id;
+  int64_t time_stamp;
+};
+
+struct AverageEntry {
+  const char* id;
+  float average_duration;
+};
+
+// Storage for keeping track of this frame's values.
+extern int num_time_logs;
+extern LogEntry time_logs[NUM_LOGS];
+
+// Storage for keeping track of average values (each entry may not be printed
+// out each frame).
+extern AverageEntry avg_entries[NUM_LOGS];
+extern int num_avg_entries;
+
+// Call this at the start of a logging phase.
+inline static void ResetTimeLog() {
+  num_time_logs = 0;
+}
+
+
+// Log a message to be printed out when printTimeLog is called, along with the
+// amount of time in ms that has passed since the last call to this function.
+inline static void TimeLog(const char* const str) {
+  LOGV("%s", str);
+  if (num_time_logs >= NUM_LOGS) {
+    LOGE("Out of log entries!");
+    return;
+  }
+
+  time_logs[num_time_logs].id = str;
+  time_logs[num_time_logs].time_stamp = CurrentThreadTimeNanos();
+  ++num_time_logs;
+}
+
+
+inline static float Blend(float old_val, float new_val) {
+  return ALPHA * old_val + (1.0f - ALPHA) * new_val;
+}
+
+
+inline static float UpdateAverage(const char* str, const float new_val) {
+  for (int entry_num = 0; entry_num < num_avg_entries; ++entry_num) {
+    AverageEntry* const entry = avg_entries + entry_num;
+    if (str == entry->id) {
+      entry->average_duration = Blend(entry->average_duration, new_val);
+      return entry->average_duration;
+    }
+  }
+
+  if (num_avg_entries >= NUM_LOGS) {
+    LOGE("Too many log entries!");
+  }
+
+  // If it wasn't there already, add it.
+  avg_entries[num_avg_entries].id = str;
+  avg_entries[num_avg_entries].average_duration = new_val;
+  ++num_avg_entries;
+
+  return new_val;
+}
+
+
+// Prints out all the timeLog statements in chronological order with the
+// interval that passed between subsequent statements.  The total time between
+// the first and last statements is printed last.
+inline static void PrintTimeLog() {
+  LogEntry* last_time = time_logs;
+
+  float average_running_total = 0.0f;
+
+  for (int i = 0; i < num_time_logs; ++i) {
+    LogEntry* const this_time = time_logs + i;
+
+    const float curr_time =
+        (this_time->time_stamp - last_time->time_stamp) / 1000000.0f;
+
+    const float avg_time = UpdateAverage(this_time->id, curr_time);
+    average_running_total += avg_time;
+
+    LOGD("%32s:    %6.3fms    %6.4fms", this_time->id, curr_time, avg_time);
+    last_time = this_time;
+  }
+
+  const float total_time =
+      (last_time->time_stamp - time_logs->time_stamp) / 1000000.0f;
+
+  LOGD("TOTAL TIME:                          %6.3fms    %6.4fms\n",
+       total_time, average_running_total);
+  LOGD(" ");
+}
+#else
+inline static void ResetTimeLog() {}
+
+inline static void TimeLog(const char* const str) {
+  LOGV("%s", str);
+}
+
+inline static void PrintTimeLog() {}
+#endif
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_TIME_LOG_H_

tensorflow/tools/android/test/jni/object_tracking/tracked_object.cc

@@ -0,0 +1,162 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/tools/android/test/jni/object_tracking/tracked_object.h"
+
+namespace tf_tracking {
+
+static const float kInitialDistance = 20.0f;
+
+static void InitNormalized(const Image<uint8_t>& src_image,
+                           const BoundingBox& position,
+                           Image<float>* const dst_image) {
+  BoundingBox scaled_box(position);
+  CopyArea(src_image, scaled_box, dst_image);
+  NormalizeImage(dst_image);
+}
+
+TrackedObject::TrackedObject(const std::string& id, const Image<uint8_t>& image,
+                             const BoundingBox& bounding_box,
+                             ObjectModelBase* const model)
+    : id_(id),
+      last_known_position_(bounding_box),
+      last_detection_position_(bounding_box),
+      position_last_computed_time_(-1),
+      object_model_(model),
+      last_detection_thumbnail_(kNormalizedThumbnailSize,
+                                kNormalizedThumbnailSize),
+      last_frame_thumbnail_(kNormalizedThumbnailSize, kNormalizedThumbnailSize),
+      tracked_correlation_(0.0f),
+      tracked_match_score_(0.0),
+      num_consecutive_frames_below_threshold_(0),
+      allowable_detection_distance_(Square(kInitialDistance)) {
+  InitNormalized(image, bounding_box, &last_detection_thumbnail_);
+}
+
+TrackedObject::~TrackedObject() {}
+
+void TrackedObject::UpdatePosition(const BoundingBox& new_position,
+                                   const int64_t timestamp,
+                                   const ImageData& image_data,
+                                   const bool authoritative) {
+  last_known_position_ = new_position;
+  position_last_computed_time_ = timestamp;
+
+  InitNormalized(*image_data.GetImage(), new_position, &last_frame_thumbnail_);
+
+  const float last_localization_correlation = ComputeCrossCorrelation(
+      last_detection_thumbnail_.data(),
+      last_frame_thumbnail_.data(),
+      last_frame_thumbnail_.data_size_);
+  LOGV("Tracked correlation to last localization:   %.6f",
+       last_localization_correlation);
+
+  // Correlation to object model, if it exists.
+  if (object_model_ != NULL) {
+    tracked_correlation_ =
+        object_model_->GetMaxCorrelation(last_frame_thumbnail_);
+    LOGV("Tracked correlation to model:               %.6f",
+         tracked_correlation_);
+
+    tracked_match_score_ =
+        object_model_->GetMatchScore(new_position, image_data);
+    LOGV("Tracked match score with model:             %.6f",
+         tracked_match_score_.value);
+  } else {
+    // If there's no model to check against, set the tracked correlation to
+    // simply be the correlation to the last set position.
+    tracked_correlation_ = last_localization_correlation;
+    tracked_match_score_ = MatchScore(0.0f);
+  }
+
+  // Determine if it's still being tracked.
+  if (tracked_correlation_ >= kMinimumCorrelationForTracking &&
+      tracked_match_score_ >= kMinimumMatchScore) {
+    num_consecutive_frames_below_threshold_ = 0;
+
+    if (object_model_ != NULL) {
+      object_model_->TrackStep(last_known_position_, *image_data.GetImage(),
+                               *image_data.GetIntegralImage(), authoritative);
+    }
+  } else if (tracked_match_score_ < kMatchScoreForImmediateTermination) {
+    if (num_consecutive_frames_below_threshold_ < 1000) {
+      LOGD("Tracked match score is way too low (%.6f), aborting track.",
+           tracked_match_score_.value);
+    }
+
+    // Add an absurd amount of missed frames so that all heuristics will
+    // consider it a lost track.
+    num_consecutive_frames_below_threshold_ += 1000;
+
+    if (object_model_ != NULL) {
+      object_model_->TrackLost();
+    }
+  } else {
+    ++num_consecutive_frames_below_threshold_;
+    allowable_detection_distance_ *= 1.1f;
+  }
+}
+
+void TrackedObject::OnDetection(ObjectModelBase* const model,
+                                const BoundingBox& detection_position,
+                                const MatchScore match_score,
+                                const int64_t timestamp,
+                                const ImageData& image_data) {
+  const float overlap = detection_position.PascalScore(last_known_position_);
+  if (overlap > kPositionOverlapThreshold) {
+    // If the position agreement with the current tracked position is good
+    // enough, lock all the current unlocked examples.
+    object_model_->TrackConfirmed();
+    num_consecutive_frames_below_threshold_ = 0;
+  }
+
+  // Before relocalizing, make sure the new proposed position is better than
+  // the existing position by a small amount to prevent thrashing.
+  if (match_score <= tracked_match_score_ + kMatchScoreBuffer) {
+    LOGI("Not relocalizing since new match is worse: %.6f < %.6f + %.6f",
+         match_score.value, tracked_match_score_.value,
+         kMatchScoreBuffer.value);
+    return;
+  }
+
+  LOGI("Relocalizing! From (%.1f, %.1f)[%.1fx%.1f] to "
+       "(%.1f, %.1f)[%.1fx%.1f]:   %.6f > %.6f",
+       last_known_position_.left_, last_known_position_.top_,
+       last_known_position_.GetWidth(), last_known_position_.GetHeight(),
+       detection_position.left_, detection_position.top_,
+       detection_position.GetWidth(), detection_position.GetHeight(),
+       match_score.value, tracked_match_score_.value);
+
+  if (overlap < kPositionOverlapThreshold) {
+    // The path might be good, it might be bad, but it's no longer a path
+    // since we're moving the box to a new position, so just nuke it from
+    // orbit to be safe.
+    object_model_->TrackLost();
+  }
+
+  object_model_ = model;
+
+  // Reset the last detected appearance.
+  InitNormalized(
+      *image_data.GetImage(), detection_position, &last_detection_thumbnail_);
+
+  num_consecutive_frames_below_threshold_ = 0;
+  last_detection_position_ = detection_position;
+
+  UpdatePosition(detection_position, timestamp, image_data, false);
+  allowable_detection_distance_ = Square(kInitialDistance);
+}
+
+}  // namespace tf_tracking

tensorflow/tools/android/test/jni/object_tracking/tracked_object.h

@@ -0,0 +1,187 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_TRACKED_OBJECT_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_TRACKED_OBJECT_H_
+
+#ifdef __RENDER_OPENGL__
+#include "tensorflow/tools/android/test/jni/object_tracking/gl_utils.h"
+#endif
+#include "tensorflow/tools/android/test/jni/object_tracking/object_detector.h"
+
+namespace tf_tracking {
+
+// A TrackedObject is a specific instance of an ObjectModel, with a known
+// position in the world.
+// It provides the last known position and number of recent detection failures,
+// in addition to the more general appearance data associated with the object
+// class (which is in ObjectModel).
+// TODO(andrewharp): Make getters/setters follow styleguide.
+class TrackedObject {
+ public:
+  TrackedObject(const std::string& id, const Image<uint8_t>& image,
+                const BoundingBox& bounding_box, ObjectModelBase* const model);
+
+  ~TrackedObject();
+
+  void UpdatePosition(const BoundingBox& new_position, const int64_t timestamp,
+                      const ImageData& image_data, const bool authoritative);
+
+  // This method is called when the tracked object is detected at a
+  // given position, and allows the associated Model to grow and/or prune
+  // itself based on where the detection occurred.
+  void OnDetection(ObjectModelBase* const model,
+                   const BoundingBox& detection_position,
+                   const MatchScore match_score, const int64_t timestamp,
+                   const ImageData& image_data);
+
+  // Called when there's no detection of the tracked object. This will cause
+  // a tracking failure after enough consecutive failures if the area under
+  // the current bounding box also doesn't meet a minimum correlation threshold
+  // with the model.
+  void OnDetectionFailure() {}
+
+  inline bool IsVisible() const {
+    return tracked_correlation_ >= kMinimumCorrelationForTracking ||
+        num_consecutive_frames_below_threshold_ < kMaxNumDetectionFailures;
+  }
+
+  inline float GetCorrelation() {
+    return tracked_correlation_;
+  }
+
+  inline MatchScore GetMatchScore() {
+    return tracked_match_score_;
+  }
+
+  inline BoundingBox GetPosition() const {
+    return last_known_position_;
+  }
+
+  inline BoundingBox GetLastDetectionPosition() const {
+    return last_detection_position_;
+  }
+
+  inline const ObjectModelBase* GetModel() const {
+    return object_model_;
+  }
+
+  inline const std::string& GetName() const {
+    return id_;
+  }
+
+  inline void Draw() const {
+#ifdef __RENDER_OPENGL__
+    if (tracked_correlation_ < kMinimumCorrelationForTracking) {
+      glColor4f(MAX(0.0f, -tracked_correlation_),
+                MAX(0.0f, tracked_correlation_),
+                0.0f,
+                1.0f);
+    } else {
+      glColor4f(MAX(0.0f, -tracked_correlation_),
+                MAX(0.0f, tracked_correlation_),
+                1.0f,
+                1.0f);
+    }
+
+    // Render the box itself.
+    BoundingBox temp_box(last_known_position_);
+    DrawBox(temp_box);
+
+    // Render a box inside this one (in case the actual box is hidden).
+    const float kBufferSize = 1.0f;
+    temp_box.left_ -= kBufferSize;
+    temp_box.top_ -= kBufferSize;
+    temp_box.right_ += kBufferSize;
+    temp_box.bottom_ += kBufferSize;
+    DrawBox(temp_box);
+
+    // Render one outside as well.
+    temp_box.left_ -= -2.0f * kBufferSize;
+    temp_box.top_ -= -2.0f * kBufferSize;
+    temp_box.right_ += -2.0f * kBufferSize;
+    temp_box.bottom_ += -2.0f * kBufferSize;
+    DrawBox(temp_box);
+#endif
+  }
+
+  // Get current object's num_consecutive_frames_below_threshold_.
+  inline int64_t GetNumConsecutiveFramesBelowThreshold() {
+    return num_consecutive_frames_below_threshold_;
+  }
+
+  // Reset num_consecutive_frames_below_threshold_ to 0.
+  inline void resetNumConsecutiveFramesBelowThreshold() {
+    num_consecutive_frames_below_threshold_ = 0;
+  }
+
+  inline float GetAllowableDistanceSquared() const {
+    return allowable_detection_distance_;
+  }
+
+ private:
+  // The unique id used throughout the system to identify this
+  // tracked object.
+  const std::string id_;
+
+  // The last known position of the object.
+  BoundingBox last_known_position_;
+
+  // The last known position of the object.
+  BoundingBox last_detection_position_;
+
+  // When the position was last computed.
+  int64_t position_last_computed_time_;
+
+  // The object model this tracked object is representative of.
+  ObjectModelBase* object_model_;
+
+  Image<float> last_detection_thumbnail_;
+
+  Image<float> last_frame_thumbnail_;
+
+  // The correlation of the object model with the preview frame at its last
+  // tracked position.
+  float tracked_correlation_;
+
+  MatchScore tracked_match_score_;
+
+  // The number of consecutive frames that the tracked position for this object
+  // has been under the correlation threshold.
+  int num_consecutive_frames_below_threshold_;
+
+  float allowable_detection_distance_;
+
+  friend std::ostream& operator<<(std::ostream& stream,
+                                  const TrackedObject& tracked_object);
+
+  TrackedObject(const TrackedObject&) = delete;
+  void operator=(const TrackedObject&) = delete;
+};
+
+inline std::ostream& operator<<(std::ostream& stream,
+                                const TrackedObject& tracked_object) {
+  stream << tracked_object.id_
+      << " " << tracked_object.last_known_position_
+      << " " << tracked_object.position_last_computed_time_
+      << " " << tracked_object.num_consecutive_frames_below_threshold_
+      << " " << tracked_object.object_model_
+      << " " << tracked_object.tracked_correlation_;
+  return stream;
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_TRACKED_OBJECT_H_

tensorflow/tools/android/test/jni/object_tracking/utils.h

@@ -0,0 +1,381 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_UTILS_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_UTILS_H_
+
+#include <math.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <time.h>
+
+#include <cmath>  // for std::abs(float)
+
+#ifndef HAVE_CLOCK_GETTIME
+// Use gettimeofday() instead of clock_gettime().
+#include <sys/time.h>
+#endif  // ifdef HAVE_CLOCK_GETTIME
+
+#include "tensorflow/tools/android/test/jni/object_tracking/logging.h"
+
+// TODO(andrewharp): clean up these macros to use the codebase statndard.
+
+// A very small number, generally used as the tolerance for accumulated
+// floating point errors in bounds-checks.
+#define EPSILON 0.00001f
+
+#define SAFE_DELETE(pointer) {\
+  if ((pointer) != NULL) {\
+    LOGV("Safe deleting pointer: %s", #pointer);\
+    delete (pointer);\
+    (pointer) = NULL;\
+  } else {\
+    LOGV("Pointer already null: %s", #pointer);\
+  }\
+}
+
+
+#ifdef __GOOGLE__
+
+#define CHECK_ALWAYS(condition, format, ...) {\
+  CHECK(condition) << StringPrintf(format, ##__VA_ARGS__);\
+}
+
+#define SCHECK(condition, format, ...) {\
+  DCHECK(condition) << StringPrintf(format, ##__VA_ARGS__);\
+}
+
+#else
+
+#define CHECK_ALWAYS(condition, format, ...) {\
+  if (!(condition)) {\
+    LOGE("CHECK FAILED (%s): " format, #condition, ##__VA_ARGS__);\
+    abort();\
+  }\
+}
+
+#ifdef SANITY_CHECKS
+#define SCHECK(condition, format, ...) {\
+  CHECK_ALWAYS(condition, format, ##__VA_ARGS__);\
+}
+#else
+#define SCHECK(condition, format, ...) {}
+#endif  // SANITY_CHECKS
+
+#endif  // __GOOGLE__
+
+
+#ifndef MAX
+#define MAX(a, b) (((a) > (b)) ? (a) : (b))
+#endif
+#ifndef MIN
+#define MIN(a, b) (((a) > (b)) ? (b) : (a))
+#endif
+
+inline static int64_t CurrentThreadTimeNanos() {
+#ifdef HAVE_CLOCK_GETTIME
+  struct timespec tm;
+  clock_gettime(CLOCK_THREAD_CPUTIME_ID, &tm);
+  return tm.tv_sec * 1000000000LL + tm.tv_nsec;
+#else
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  return tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
+#endif
+}
+
+inline static int64_t CurrentRealTimeMillis() {
+#ifdef HAVE_CLOCK_GETTIME
+  struct timespec tm;
+  clock_gettime(CLOCK_MONOTONIC, &tm);
+  return tm.tv_sec * 1000LL + tm.tv_nsec / 1000000LL;
+#else
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  return tv.tv_sec * 1000 + tv.tv_usec / 1000;
+#endif
+}
+
+
+template<typename T>
+inline static T Square(const T a) {
+  return a * a;
+}
+
+
+template<typename T>
+inline static T Clip(const T a, const T floor, const T ceil) {
+  SCHECK(ceil >= floor, "Bounds mismatch!");
+  return (a <= floor) ? floor : ((a >= ceil) ? ceil : a);
+}
+
+
+template<typename T>
+inline static int Floor(const T a) {
+  return static_cast<int>(a);
+}
+
+
+template<typename T>
+inline static int Ceil(const T a) {
+  return Floor(a) + 1;
+}
+
+
+template<typename T>
+inline static bool InRange(const T a, const T min, const T max) {
+  return (a >= min) && (a <= max);
+}
+
+
+inline static bool ValidIndex(const int a, const int max) {
+  return (a >= 0) && (a < max);
+}
+
+
+inline bool NearlyEqual(const float a, const float b, const float tolerance) {
+  return std::abs(a - b) < tolerance;
+}
+
+
+inline bool NearlyEqual(const float a, const float b) {
+  return NearlyEqual(a, b, EPSILON);
+}
+
+
+template<typename T>
+inline static int Round(const float a) {
+  return (a - static_cast<float>(floor(a) > 0.5f) ? ceil(a) : floor(a));
+}
+
+
+template<typename T>
+inline static void Swap(T* const a, T* const b) {
+  // Cache out the VALUE of what's at a.
+  T tmp = *a;
+  *a = *b;
+
+  *b = tmp;
+}
+
+
+static inline float randf() {
+  return rand() / static_cast<float>(RAND_MAX);
+}
+
+static inline float randf(const float min_value, const float max_value) {
+  return randf() * (max_value - min_value) + min_value;
+}
+
+static inline uint16_t RealToFixed115(const float real_number) {
+  SCHECK(InRange(real_number, 0.0f, 2048.0f),
+        "Value out of range! %.2f", real_number);
+
+  static const float kMult = 32.0f;
+  const float round_add = (real_number > 0.0f) ? 0.5f : -0.5f;
+  return static_cast<uint16_t>(real_number * kMult + round_add);
+}
+
+static inline float FixedToFloat115(const uint16_t fp_number) {
+  const float kDiv = 32.0f;
+  return (static_cast<float>(fp_number) / kDiv);
+}
+
+static inline int RealToFixed1616(const float real_number) {
+  static const float kMult = 65536.0f;
+  SCHECK(InRange(real_number, -kMult, kMult),
+        "Value out of range! %.2f", real_number);
+
+  const float round_add = (real_number > 0.0f) ? 0.5f : -0.5f;
+  return static_cast<int>(real_number * kMult + round_add);
+}
+
+static inline float FixedToFloat1616(const int fp_number) {
+  const float kDiv = 65536.0f;
+  return (static_cast<float>(fp_number) / kDiv);
+}
+
+template<typename T>
+// produces numbers in range [0,2*M_PI] (rather than -PI,PI)
+inline T FastAtan2(const T y, const T x) {
+  static const T coeff_1 = (T)(M_PI / 4.0);
+  static const T coeff_2 = (T)(3.0 * coeff_1);
+  const T abs_y = fabs(y);
+  T angle;
+  if (x >= 0) {
+    T r = (x - abs_y) / (x + abs_y);
+    angle = coeff_1 - coeff_1 * r;
+  } else {
+    T r = (x + abs_y) / (abs_y - x);
+    angle = coeff_2 - coeff_1 * r;
+  }
+  static const T PI_2 = 2.0 * M_PI;
+  return y < 0 ? PI_2 - angle : angle;
+}
+
+#define NELEMS(X) (sizeof(X) / sizeof(X[0]))
+
+namespace tf_tracking {
+
+#ifdef __ARM_NEON
+float ComputeMeanNeon(const float* const values, const int num_vals);
+
+float ComputeStdDevNeon(const float* const values, const int num_vals,
+                        const float mean);
+
+float ComputeWeightedMeanNeon(const float* const values,
+                              const float* const weights, const int num_vals);
+
+float ComputeCrossCorrelationNeon(const float* const values1,
+                                  const float* const values2,
+                                  const int num_vals);
+#endif
+
+inline float ComputeMeanCpu(const float* const values, const int num_vals) {
+  // Get mean.
+  float sum = values[0];
+  for (int i = 1; i < num_vals; ++i) {
+    sum += values[i];
+  }
+  return sum / static_cast<float>(num_vals);
+}
+
+
+inline float ComputeMean(const float* const values, const int num_vals) {
+  return
+#ifdef __ARM_NEON
+      (num_vals >= 8) ? ComputeMeanNeon(values, num_vals) :
+#endif
+                      ComputeMeanCpu(values, num_vals);
+}
+
+
+inline float ComputeStdDevCpu(const float* const values,
+                              const int num_vals,
+                              const float mean) {
+  // Get Std dev.
+  float squared_sum = 0.0f;
+  for (int i = 0; i < num_vals; ++i) {
+    squared_sum += Square(values[i] - mean);
+  }
+  return sqrt(squared_sum / static_cast<float>(num_vals));
+}
+
+
+inline float ComputeStdDev(const float* const values,
+                           const int num_vals,
+                           const float mean) {
+  return
+#ifdef __ARM_NEON
+      (num_vals >= 8) ? ComputeStdDevNeon(values, num_vals, mean) :
+#endif
+                      ComputeStdDevCpu(values, num_vals, mean);
+}
+
+
+// TODO(andrewharp): Accelerate with NEON.
+inline float ComputeWeightedMean(const float* const values,
+                                 const float* const weights,
+                                 const int num_vals) {
+  float sum = 0.0f;
+  float total_weight = 0.0f;
+  for (int i = 0; i < num_vals; ++i) {
+    sum += values[i] * weights[i];
+    total_weight += weights[i];
+  }
+  return sum / num_vals;
+}
+
+
+inline float ComputeCrossCorrelationCpu(const float* const values1,
+                                        const float* const values2,
+                                        const int num_vals) {
+  float sxy = 0.0f;
+  for (int offset = 0; offset < num_vals; ++offset) {
+    sxy += values1[offset] * values2[offset];
+  }
+
+  const float cross_correlation = sxy / num_vals;
+
+  return cross_correlation;
+}
+
+
+inline float ComputeCrossCorrelation(const float* const values1,
+                                     const float* const values2,
+                                     const int num_vals) {
+  return
+#ifdef __ARM_NEON
+      (num_vals >= 8) ? ComputeCrossCorrelationNeon(values1, values2, num_vals)
+                      :
+#endif
+                      ComputeCrossCorrelationCpu(values1, values2, num_vals);
+}
+
+
+inline void NormalizeNumbers(float* const values, const int num_vals) {
+  // Find the mean and then subtract so that the new mean is 0.0.
+  const float mean = ComputeMean(values, num_vals);
+  VLOG(2) << "Mean is " << mean;
+  float* curr_data = values;
+  for (int i = 0; i < num_vals; ++i) {
+    *curr_data -= mean;
+    curr_data++;
+  }
+
+  // Now divide by the std deviation so the new standard deviation is 1.0.
+  // The numbers might all be identical (and thus shifted to 0.0 now),
+  // so only scale by the standard deviation if this is not the case.
+  const float std_dev = ComputeStdDev(values, num_vals, 0.0f);
+  if (std_dev > 0.0f) {
+    VLOG(2) << "Std dev is " << std_dev;
+    curr_data = values;
+    for (int i = 0; i < num_vals; ++i) {
+      *curr_data /= std_dev;
+      curr_data++;
+    }
+  }
+}
+
+
+// Returns the determinant of a 2x2 matrix.
+template<class T>
+inline T FindDeterminant2x2(const T* const a) {
+  // Determinant: (ad - bc)
+  return a[0] * a[3] - a[1] * a[2];
+}
+
+
+// Finds the inverse of a 2x2 matrix.
+// Returns true upon success, false if the matrix is not invertible.
+template<class T>
+inline bool Invert2x2(const T* const a, float* const a_inv) {
+  const float det = static_cast<float>(FindDeterminant2x2(a));
+  if (fabs(det) < EPSILON) {
+    return false;
+  }
+  const float inv_det = 1.0f / det;
+
+  a_inv[0] = inv_det * static_cast<float>(a[3]);   // d
+  a_inv[1] = inv_det * static_cast<float>(-a[1]);  // -b
+  a_inv[2] = inv_det * static_cast<float>(-a[2]);  // -c
+  a_inv[3] = inv_det * static_cast<float>(a[0]);   // a
+
+  return true;
+}
+
+}  // namespace tf_tracking
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_OBJECT_TRACKING_UTILS_H_

tensorflow/tools/android/test/jni/object_tracking/utils_neon.cc

@@ -0,0 +1,151 @@
+/* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// NEON implementations of Image methods for compatible devices.  Control
+// should never enter this compilation unit on incompatible devices.
+
+#ifdef __ARM_NEON
+
+#include <arm_neon.h>
+
+#include "tensorflow/tools/android/test/jni/object_tracking/geom.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image-inl.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/image.h"
+#include "tensorflow/tools/android/test/jni/object_tracking/utils.h"
+
+namespace tf_tracking {
+
+inline static float GetSum(const float32x4_t& values) {
+  static float32_t summed_values[4];
+  vst1q_f32(summed_values, values);
+  return summed_values[0]
+       + summed_values[1]
+       + summed_values[2]
+       + summed_values[3];
+}
+
+
+float ComputeMeanNeon(const float* const values, const int num_vals) {
+  SCHECK(num_vals >= 8, "Not enough values to merit NEON: %d", num_vals);
+
+  const float32_t* const arm_vals = (const float32_t* const) values;
+  float32x4_t accum = vdupq_n_f32(0.0f);
+
+  int offset = 0;
+  for (; offset <= num_vals - 4; offset += 4) {
+    accum = vaddq_f32(accum, vld1q_f32(&arm_vals[offset]));
+  }
+
+  // Pull the accumulated values into a single variable.
+  float sum = GetSum(accum);
+
+  // Get the remaining 1 to 3 values.
+  for (; offset < num_vals; ++offset) {
+    sum += values[offset];
+  }
+
+  const float mean_neon = sum / static_cast<float>(num_vals);
+
+#ifdef SANITY_CHECKS
+  const float mean_cpu = ComputeMeanCpu(values, num_vals);
+  SCHECK(NearlyEqual(mean_neon, mean_cpu, EPSILON * num_vals),
+        "Neon mismatch with CPU mean! %.10f vs %.10f",
+        mean_neon, mean_cpu);
+#endif
+
+  return mean_neon;
+}
+
+
+float ComputeStdDevNeon(const float* const values,
+                        const int num_vals, const float mean) {
+  SCHECK(num_vals >= 8, "Not enough values to merit NEON: %d", num_vals);
+
+  const float32_t* const arm_vals = (const float32_t* const) values;
+  const float32x4_t mean_vec = vdupq_n_f32(-mean);
+
+  float32x4_t accum = vdupq_n_f32(0.0f);
+
+  int offset = 0;
+  for (; offset <= num_vals - 4; offset += 4) {
+    const float32x4_t deltas =
+        vaddq_f32(mean_vec, vld1q_f32(&arm_vals[offset]));
+
+    accum = vmlaq_f32(accum, deltas, deltas);
+  }
+
+  // Pull the accumulated values into a single variable.
+  float squared_sum = GetSum(accum);
+
+  // Get the remaining 1 to 3 values.
+  for (; offset < num_vals; ++offset) {
+    squared_sum += Square(values[offset] - mean);
+  }
+
+  const float std_dev_neon = sqrt(squared_sum / static_cast<float>(num_vals));
+
+#ifdef SANITY_CHECKS
+  const float std_dev_cpu = ComputeStdDevCpu(values, num_vals, mean);
+  SCHECK(NearlyEqual(std_dev_neon, std_dev_cpu, EPSILON * num_vals),
+        "Neon mismatch with CPU std dev! %.10f vs %.10f",
+        std_dev_neon, std_dev_cpu);
+#endif
+
+  return std_dev_neon;
+}
+
+
+float ComputeCrossCorrelationNeon(const float* const values1,
+                                  const float* const values2,
+                                  const int num_vals) {
+  SCHECK(num_vals >= 8, "Not enough values to merit NEON: %d", num_vals);
+
+  const float32_t* const arm_vals1 = (const float32_t* const) values1;
+  const float32_t* const arm_vals2 = (const float32_t* const) values2;
+
+  float32x4_t accum = vdupq_n_f32(0.0f);
+
+  int offset = 0;
+  for (; offset <= num_vals - 4; offset += 4) {
+    accum = vmlaq_f32(accum,
+                      vld1q_f32(&arm_vals1[offset]),
+                      vld1q_f32(&arm_vals2[offset]));
+  }
+
+  // Pull the accumulated values into a single variable.
+  float sxy = GetSum(accum);
+
+  // Get the remaining 1 to 3 values.
+  for (; offset < num_vals; ++offset) {
+    sxy += values1[offset] * values2[offset];
+  }
+
+  const float cross_correlation_neon = sxy / num_vals;
+
+#ifdef SANITY_CHECKS
+  const float cross_correlation_cpu =
+      ComputeCrossCorrelationCpu(values1, values2, num_vals);
+  SCHECK(NearlyEqual(cross_correlation_neon, cross_correlation_cpu,
+                    EPSILON * num_vals),
+        "Neon mismatch with CPU cross correlation! %.10f vs %.10f",
+        cross_correlation_neon, cross_correlation_cpu);
+#endif
+
+  return cross_correlation_neon;
+}
+
+}  // namespace tf_tracking
+
+#endif  // __ARM_NEON

tensorflow/tools/android/test/jni/rgb2yuv.cc

@@ -0,0 +1,99 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// These utility functions allow for the conversion of RGB data to YUV data.
+
+#include "tensorflow/tools/android/test/jni/rgb2yuv.h"
+
+static inline void WriteYUV(const int x, const int y, const int width,
+                            const int r8, const int g8, const int b8,
+                            uint8_t* const pY, uint8_t* const pUV) {
+  // Using formulas from http://msdn.microsoft.com/en-us/library/ms893078
+  *pY = ((66 * r8 + 129 * g8 + 25 * b8 + 128) >> 8) + 16;
+
+  // Odd widths get rounded up so that UV blocks on the side don't get cut off.
+  const int blocks_per_row = (width + 1) / 2;
+
+  // 2 bytes per UV block
+  const int offset = 2 * (((y / 2) * blocks_per_row + (x / 2)));
+
+  // U and V are the average values of all 4 pixels in the block.
+  if (!(x & 1) && !(y & 1)) {
+    // Explicitly clear the block if this is the first pixel in it.
+    pUV[offset] = 0;
+    pUV[offset + 1] = 0;
+  }
+
+  // V (with divide by 4 factored in)
+#ifdef __APPLE__
+  const int u_offset = 0;
+  const int v_offset = 1;
+#else
+  const int u_offset = 1;
+  const int v_offset = 0;
+#endif
+  pUV[offset + v_offset] += ((112 * r8 - 94 * g8 - 18 * b8 + 128) >> 10) + 32;
+
+  // U (with divide by 4 factored in)
+  pUV[offset + u_offset] += ((-38 * r8 - 74 * g8 + 112 * b8 + 128) >> 10) + 32;
+}
+
+void ConvertARGB8888ToYUV420SP(const uint32_t* const input,
+                               uint8_t* const output, int width, int height) {
+  uint8_t* pY = output;
+  uint8_t* pUV = output + (width * height);
+  const uint32_t* in = input;
+
+  for (int y = 0; y < height; y++) {
+    for (int x = 0; x < width; x++) {
+      const uint32_t rgb = *in++;
+#ifdef __APPLE__
+      const int nB = (rgb >> 8) & 0xFF;
+      const int nG = (rgb >> 16) & 0xFF;
+      const int nR = (rgb >> 24) & 0xFF;
+#else
+      const int nR = (rgb >> 16) & 0xFF;
+      const int nG = (rgb >> 8) & 0xFF;
+      const int nB = rgb & 0xFF;
+#endif
+      WriteYUV(x, y, width, nR, nG, nB, pY++, pUV);
+    }
+  }
+}
+
+void ConvertRGB565ToYUV420SP(const uint16_t* const input, uint8_t* const output,
+                             const int width, const int height) {
+  uint8_t* pY = output;
+  uint8_t* pUV = output + (width * height);
+  const uint16_t* in = input;
+
+  for (int y = 0; y < height; y++) {
+    for (int x = 0; x < width; x++) {
+      const uint32_t rgb = *in++;
+
+      const int r5 = ((rgb >> 11) & 0x1F);
+      const int g6 = ((rgb >> 5) & 0x3F);
+      const int b5 = (rgb & 0x1F);
+
+      // Shift left, then fill in the empty low bits with a copy of the high
+      // bits so we can stretch across the entire 0 - 255 range.
+      const int r8 = r5 << 3 | r5 >> 2;
+      const int g8 = g6 << 2 | g6 >> 4;
+      const int b8 = b5 << 3 | b5 >> 2;
+
+      WriteYUV(x, y, width, r8, g8, b8, pY++, pUV);
+    }
+  }
+}

tensorflow/tools/android/test/jni/rgb2yuv.h

@@ -0,0 +1,35 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_RGB2YUV_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_RGB2YUV_H_
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ConvertARGB8888ToYUV420SP(const uint32_t* const input,
+                               uint8_t* const output, int width, int height);
+
+void ConvertRGB565ToYUV420SP(const uint16_t* const input, uint8_t* const output,
+                             const int width, const int height);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_RGB2YUV_H_

tensorflow/tools/android/test/jni/version_script.lds

@@ -0,0 +1,11 @@
+VERS_1.0 {
+  # Export JNI symbols.
+  global:
+    Java_*;
+    JNI_OnLoad;
+    JNI_OnUnload;
+
+  # Hide everything else.
+  local:
+    *;
+};

tensorflow/tools/android/test/jni/yuv2rgb.cc

@@ -0,0 +1,191 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This is a collection of routines which converts various YUV image formats
+// to ARGB.
+
+#include "tensorflow/tools/android/test/jni/yuv2rgb.h"
+
+#ifndef MAX
+#define MAX(a, b) ({__typeof__(a) _a = (a); __typeof__(b) _b = (b); _a > _b ? _a : _b; })
+#define MIN(a, b) ({__typeof__(a) _a = (a); __typeof__(b) _b = (b); _a < _b ? _a : _b; })
+#endif
+
+// This value is 2 ^ 18 - 1, and is used to clamp the RGB values before their ranges
+// are normalized to eight bits.
+static const int kMaxChannelValue = 262143;
+
+static inline uint32_t YUV2RGB(int nY, int nU, int nV) {
+  nY -= 16;
+  nU -= 128;
+  nV -= 128;
+  if (nY < 0) nY = 0;
+
+  // This is the floating point equivalent. We do the conversion in integer
+  // because some Android devices do not have floating point in hardware.
+  // nR = (int)(1.164 * nY + 2.018 * nU);
+  // nG = (int)(1.164 * nY - 0.813 * nV - 0.391 * nU);
+  // nB = (int)(1.164 * nY + 1.596 * nV);
+
+  int nR = 1192 * nY + 1634 * nV;
+  int nG = 1192 * nY - 833 * nV - 400 * nU;
+  int nB = 1192 * nY + 2066 * nU;
+
+  nR = MIN(kMaxChannelValue, MAX(0, nR));
+  nG = MIN(kMaxChannelValue, MAX(0, nG));
+  nB = MIN(kMaxChannelValue, MAX(0, nB));
+
+  nR = (nR >> 10) & 0xff;
+  nG = (nG >> 10) & 0xff;
+  nB = (nB >> 10) & 0xff;
+
+  return 0xff000000 | (nR << 16) | (nG << 8) | nB;
+}
+
+//  Accepts a YUV 4:2:0 image with a plane of 8 bit Y samples followed by
+//  separate u and v planes with arbitrary row and column strides,
+//  containing 8 bit 2x2 subsampled chroma samples.
+//  Converts to a packed ARGB 32 bit output of the same pixel dimensions.
+void ConvertYUV420ToARGB8888(const uint8_t* const yData,
+                             const uint8_t* const uData,
+                             const uint8_t* const vData, uint32_t* const output,
+                             const int width, const int height,
+                             const int y_row_stride, const int uv_row_stride,
+                             const int uv_pixel_stride) {
+  uint32_t* out = output;
+
+  for (int y = 0; y < height; y++) {
+    const uint8_t* pY = yData + y_row_stride * y;
+
+    const int uv_row_start = uv_row_stride * (y >> 1);
+    const uint8_t* pU = uData + uv_row_start;
+    const uint8_t* pV = vData + uv_row_start;
+
+    for (int x = 0; x < width; x++) {
+      const int uv_offset = (x >> 1) * uv_pixel_stride;
+      *out++ = YUV2RGB(pY[x], pU[uv_offset], pV[uv_offset]);
+    }
+  }
+}
+
+//  Accepts a YUV 4:2:0 image with a plane of 8 bit Y samples followed by an
+//  interleaved U/V plane containing 8 bit 2x2 subsampled chroma samples,
+//  except the interleave order of U and V is reversed. Converts to a packed
+//  ARGB 32 bit output of the same pixel dimensions.
+void ConvertYUV420SPToARGB8888(const uint8_t* const yData,
+                               const uint8_t* const uvData,
+                               uint32_t* const output, const int width,
+                               const int height) {
+  const uint8_t* pY = yData;
+  const uint8_t* pUV = uvData;
+  uint32_t* out = output;
+
+  for (int y = 0; y < height; y++) {
+    for (int x = 0; x < width; x++) {
+      int nY = *pY++;
+      int offset = (y >> 1) * width + 2 * (x >> 1);
+#ifdef __APPLE__
+      int nU = pUV[offset];
+      int nV = pUV[offset + 1];
+#else
+      int nV = pUV[offset];
+      int nU = pUV[offset + 1];
+#endif
+
+      *out++ = YUV2RGB(nY, nU, nV);
+    }
+  }
+}
+
+// The same as above, but downsamples each dimension to half size.
+void ConvertYUV420SPToARGB8888HalfSize(const uint8_t* const input,
+                                       uint32_t* const output, int width,
+                                       int height) {
+  const uint8_t* pY = input;
+  const uint8_t* pUV = input + (width * height);
+  uint32_t* out = output;
+  int stride = width;
+  width >>= 1;
+  height >>= 1;
+
+  for (int y = 0; y < height; y++) {
+    for (int x = 0; x < width; x++) {
+      int nY = (pY[0] + pY[1] + pY[stride] + pY[stride + 1]) >> 2;
+      pY += 2;
+#ifdef __APPLE__
+      int nU = *pUV++;
+      int nV = *pUV++;
+#else
+      int nV = *pUV++;
+      int nU = *pUV++;
+#endif
+
+      *out++ = YUV2RGB(nY, nU, nV);
+    }
+    pY += stride;
+  }
+}
+
+//  Accepts a YUV 4:2:0 image with a plane of 8 bit Y samples followed by an
+//  interleaved U/V plane containing 8 bit 2x2 subsampled chroma samples,
+//  except the interleave order of U and V is reversed. Converts to a packed
+//  RGB 565 bit output of the same pixel dimensions.
+void ConvertYUV420SPToRGB565(const uint8_t* const input, uint16_t* const output,
+                             const int width, const int height) {
+  const uint8_t* pY = input;
+  const uint8_t* pUV = input + (width * height);
+  uint16_t* out = output;
+
+  for (int y = 0; y < height; y++) {
+    for (int x = 0; x < width; x++) {
+      int nY = *pY++;
+      int offset = (y >> 1) * width + 2 * (x >> 1);
+#ifdef __APPLE__
+      int nU = pUV[offset];
+      int nV = pUV[offset + 1];
+#else
+      int nV = pUV[offset];
+      int nU = pUV[offset + 1];
+#endif
+
+      nY -= 16;
+      nU -= 128;
+      nV -= 128;
+      if (nY < 0) nY = 0;
+
+      // This is the floating point equivalent. We do the conversion in integer
+      // because some Android devices do not have floating point in hardware.
+      // nR = (int)(1.164 * nY + 2.018 * nU);
+      // nG = (int)(1.164 * nY - 0.813 * nV - 0.391 * nU);
+      // nB = (int)(1.164 * nY + 1.596 * nV);
+
+      int nR = 1192 * nY + 1634 * nV;
+      int nG = 1192 * nY - 833 * nV - 400 * nU;
+      int nB = 1192 * nY + 2066 * nU;
+
+      nR = MIN(kMaxChannelValue, MAX(0, nR));
+      nG = MIN(kMaxChannelValue, MAX(0, nG));
+      nB = MIN(kMaxChannelValue, MAX(0, nB));
+
+      // Shift more than for ARGB8888 and apply appropriate bitmask.
+      nR = (nR >> 13) & 0x1f;
+      nG = (nG >> 12) & 0x3f;
+      nB = (nB >> 13) & 0x1f;
+
+      // R is high 5 bits, G is middle 6 bits, and B is low 5 bits.
+      *out++ = (nR << 11) | (nG << 5) | nB;
+    }
+  }
+}

tensorflow/tools/android/test/jni/yuv2rgb.h

@@ -0,0 +1,57 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This is a collection of routines which converts various YUV image formats
+// to (A)RGB.
+
+#ifndef TENSORFLOW_TOOLS_ANDROID_TEST_JNI_YUV2RGB_H_
+#define TENSORFLOW_TOOLS_ANDROID_TEST_JNI_YUV2RGB_H_
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ConvertYUV420ToARGB8888(const uint8_t* const yData,
+                             const uint8_t* const uData,
+                             const uint8_t* const vData, uint32_t* const output,
+                             const int width, const int height,
+                             const int y_row_stride, const int uv_row_stride,
+                             const int uv_pixel_stride);
+
+// Converts YUV420 semi-planar data to ARGB 8888 data using the supplied width
+// and height. The input and output must already be allocated and non-null.
+// For efficiency, no error checking is performed.
+void ConvertYUV420SPToARGB8888(const uint8_t* const pY,
+                               const uint8_t* const pUV, uint32_t* const output,
+                               const int width, const int height);
+
+// The same as above, but downsamples each dimension to half size.
+void ConvertYUV420SPToARGB8888HalfSize(const uint8_t* const input,
+                                       uint32_t* const output, int width,
+                                       int height);
+
+// Converts YUV420 semi-planar data to RGB 565 data using the supplied width
+// and height. The input and output must already be allocated and non-null.
+// For efficiency, no error checking is performed.
+void ConvertYUV420SPToRGB565(const uint8_t* const input, uint16_t* const output,
+                             const int width, const int height);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // TENSORFLOW_TOOLS_ANDROID_TEST_JNI_YUV2RGB_H_