---
name: ai-tracking-box
description: Animated bounding box with L-shaped corner markers following an oscillating path — simulates AI object detection / tracking.
metadata:
tags: ai, tracking, bounding-box, detection, corner, yellow, ml
---
# AI Tracking Box
A bounding box with corner markers ("L-brackets") that follows a moving target, simulating real-time AI detection. Position and size oscillate on sine paths to mimic continuous re-computation. Conventionally rendered in "AI detection yellow" (`{detectionYellow}`) on a dark background with a confidence label.
## How It Works
- Box position `(x, y)` and size `(w, h)` are derived from sine + drift across composition time
- 4 L-bracket corner markers (`
` per corner with two-sided borders) sit ON the box
- Optional label tag above the top-left corner showing class name + confidence percent
All driven by GSAP timeline so HF seeks deterministically.
## HTML
```html
{targetGlyph} {LABEL} · {confidence}%
```
## CSS
```css
.scene {
position: relative;
width: 100%;
height: 100%;
background: radial-gradient(ellipse at center, {bgInner} 0%, {bgOuter} 70%);
font-family: {font};
overflow: hidden;
}
.bg {
position: absolute;
inset: 0;
display: grid;
place-items: center;
gap: 60px;
}
.bg-content {
position: absolute;
top: 120px;
left: 50%;
transform: translateX(-50%);
font-size: 80px;
font-weight: 900;
color: {bgTextColor};
letter-spacing: 12px;
text-transform: uppercase;
}
.bg-mascot {
position: absolute;
font-size: 240px;
line-height: 1;
}
.track-box {
position: absolute;
/* Position + size set by GSAP onUpdate */
pointer-events: none;
will-change: transform, width, height;
}
.corner {
position: absolute;
width: 48px;
height: 48px;
}
.corner.tl {
top: -8px;
left: -8px;
border-top: 6px solid {detectionYellow};
border-left: 6px solid {detectionYellow};
}
.corner.tr {
top: -8px;
right: -8px;
border-top: 6px solid {detectionYellow};
border-right: 6px solid {detectionYellow};
}
.corner.bl {
bottom: -8px;
left: -8px;
border-bottom: 6px solid {detectionYellow};
border-left: 6px solid {detectionYellow};
}
.corner.br {
bottom: -8px;
right: -8px;
border-bottom: 6px solid {detectionYellow};
border-right: 6px solid {detectionYellow};
}
.label {
position: absolute;
top: -56px;
left: -8px;
padding: 8px 16px;
background: {detectionYellow};
color: {labelTextColor};
font-family: {monoFont};
font-size: 24px;
font-weight: 800;
letter-spacing: 2px;
border-radius: 6px;
white-space: nowrap;
}
```
## GSAP Timeline
```html
```
## How to Choose Values
- **ENTRY_SCALE** — start scale of the box before it pops in
- Range: 0.5-0.9
- Effects: low end = stronger pop / more "snapping into focus"; high end = subtle reveal
- Constraints: must be < 1 (box scales UP into place)
- Reference: examples use 0.7
- **ENTRY_DUR** — duration of the fade-in + scale-up (seconds)
- Range: 0.3-0.8 s
- Effects: low end = snappy / authoritative lock-on; high end = soft / observational
- Constraints: should end before TRACK_START
- Reference: examples use 0.5
- **ENTRY_START** — when the entry tween begins (seconds, absolute on timeline)
- Range: 0-2 s typically
- Effects: late start = lets the viewer notice the target first before the AI "finds" it; early start = AI is already watching
- Constraints: ENTRY_START + ENTRY_DUR ≤ TRACK_START
- Reference: examples use 0.5
- **ENTRY_BOUNCE** — coefficient passed to `back.out(...)` on entry
- Range: 1.2-2.5
- Effects: low end = subtle overshoot; high end = exaggerated snap (reads as "aggressive lock-on")
- Constraints: stick to `back.out` family — `elastic` reads as cartoonish, `power` reads as flat
- Reference: examples use 1.4
- **TRACK_START** — when continuous tracking begins (seconds)
- Range: ≥ ENTRY_START + ENTRY_DUR
- Effects: gap between entry and tracking = pause for emphasis; no gap = seamless lock + follow
- Constraints: TRACK_START + TRACK_DUR ≤ composition duration
- Reference: examples use 1.0
- **TRACK_DUR** — length of the tracking phase (seconds)
- Range: 2-8 s
- Effects: short = "quick scan"; long = "sustained observation"
- Constraints: must accommodate at least one full CYCLE to read as oscillation
- Reference: examples use 4.0
- **CYCLES** — number of full sine oscillations of the target across TRACK_DUR
- Range: 0.5-3
- Effects: low = lazy drift; high = jittery / hyperactive target
- Constraints: CYCLES / TRACK_DUR sets effective Hz of drift; keep < ~0.6 Hz or motion blurs
- Reference: examples use 1.5
- **DRIFT_X / DRIFT_Y** — amplitude of target oscillation around screen center (px, 1920×1080 basis)
- Range: 40-200 px
- Effects: small = subtle hover; large = wide chase that pushes the box near the frame edge
- Constraints: SCREEN_CENTER ± DRIFT must keep mascot fully on screen given MASCOT_SIZE
- Reference: examples use 80 / 50
- **SIZE_BASE** — mean width/height of the bounding box (px)
- Range: 200-500 px
- Effects: small = "specimen tag"; large = "the box IS the subject"
- Constraints: must visibly enclose the target glyph at all confidence sizes
- Reference: examples use 320
- **SIZE_VAR** — half-amplitude of per-frame size jitter (px)
- Range: 5-10% of SIZE_BASE
- Effects: low end = stable / confident detector; high end = jittery / re-fitting detector. Outside this range reads as "broken" (too much) or "static UI" (none)
- Constraints: keep < 0.15 × SIZE_BASE to avoid breaking the L-bracket illusion
- Reference: examples use 30 (~9% of 320)
- **SIZE_FREQ_MULT** — multiplier on tracking phase for size oscillation
- Range: 1.5-3
- Effects: 1 = size pulses in lock-step with drift (reads mechanical); irrational ratio = organic re-fitting
- Constraints: avoid integer ratios; non-integer reads as continuous recomputation
- Reference: examples use 2.3
- **MASCOT_SIZE** — rendered width of the mascot element (px); used to center it on (mx, my)
- Range: matches CSS `font-size` of `.bg-mascot`
- Effects: must match the actual rendered size or mascot drifts out of the box
- Constraints: MASCOT_SIZE / 2 = offset applied to top/left
- Reference: examples use 240 (matches `font-size: 240px`)
- **CONFIDENCE_MEAN** — center % shown on the label
- Range: 95-99
- Effects: < 95 reads "uncertain"; 100 reads "fake-precise". 97 is the sweet spot for "confident AI"
- Constraints: keep CONFIDENCE_MEAN + CONFIDENCE_VAR ≤ 99
- Reference: examples use 97
- **CONFIDENCE_VAR** — flicker half-range around CONFIDENCE_MEAN
- Range: 1-3
- Effects: 0 = static (looks like a screenshot); >3 = unstable (looks broken)
- Constraints: CONFIDENCE_MEAN ± CONFIDENCE_VAR ⊂ [95, 99]
- Reference: examples use 2
- **CONFIDENCE_FREQ_MULT** — multiplier on tracking phase for label flicker
- Range: 3-6
- Effects: low = synced with drift (mechanical); high = fast nervous flicker (reads "live inference")
- Constraints: keep above SIZE_FREQ_MULT so label flickers faster than the box breathes
- Reference: examples use 4
- **COMP_WIDTH / COMP_HEIGHT** — composition pixel dimensions (used to derive SCREEN_CENTER)
- Range: dictated by the HF composition (`data-width` / `data-height`)
- Effects: not a creative choice — match the parent composition
- Constraints: SCREEN_CENTER = (COMP_WIDTH/2, COMP_HEIGHT/2)
- Reference: examples use 1920 × 1080
- **{detectionYellow}** — corner-marker + label background color
- This is a **discrete convention**, not a tunable range. AI detection overlays are yellow on dark backgrounds across the industry (autonomous-vehicle HUDs, security CV, ML demos). Red reads as "warning", green as "success", blue as "info" — none read as "detection."
- Recommended: a saturated warm yellow (`#facc15` / `#FCD34D` family) on a dark navy or near-black background. Substituting any other hue loses genre legibility.
- Reference: examples use `#facc15` (Tailwind `yellow-400`)
- **{bgInner} / {bgOuter}** — radial-gradient background stops
- Should be dark and low-chroma so the yellow markers pop
- Constraints: choose colors with sufficient contrast against {detectionYellow} (the corners and label must remain readable)
- Reference: examples use `#161a3a` (inner) → `#0b0d1f` (outer)
- **{labelTextColor}** — text color inside the yellow label tag
- Constraints: must contrast against {detectionYellow}; typically the same near-black as {bgOuter}
- Reference: examples use `#0b0d1f`
- **{font} / {monoFont}** — scene text font and label font
- {font}: sans-serif body font for {Brand} backdrop
- {monoFont}: monospaced font for the confidence label (mono reinforces "machine readout" affordance)
- Reference: examples use `"Inter", sans-serif` and `"JetBrains Mono", monospace`
## Variations
### Multi-object detection
Multiple boxes at different phases (each tracking its own mascot). Each is its own onUpdate-driven set; offset their phase by `Math.PI / N` so they don't tick synchronously.
### Lost-then-reacquired
The box fades to {LOST_OPACITY} (~{LOST_DUR}) then re-snaps to a new position with a "REACQUIRED" label flash:
```js
tl.to(box, { opacity: LOST_OPACITY, duration: LOST_DUR }, LOST_START);
tl.to(
box,
{ opacity: 1.0, duration: REACQUIRE_DUR, ease: `back.out(${REACQUIRE_BOUNCE})` },
REACQUIRE_START,
);
tl.to(label, { textContent: "REACQUIRED · 99%", duration: 0 }, REACQUIRE_START);
```
(LOST_OPACITY ≈ 0.2-0.4; REACQUIRE_BOUNCE ≈ 1.8-2.5 for a snappier re-lock than the initial entry.)
### Tracking-then-zoom
After tracking, the camera (via [viewport-change](viewport-change.md)) zooms into the tracked box. Combined effect: "the AI found something, now show it."
## Key Principles
- **Yellow on dark background is the detection convention** — see {detectionYellow} entry. Other colors lose the genre signal.
- **Box ALWAYS contains the target** — recompute box position EVERY frame from target position; never trail behind. If the box lags, it reads as "broken tracker," not "smart AI."
- **Subtle size variation (~5-10% of SIZE_BASE)** — too much and the tracker looks confused; just right reads as "real-time recomputation."
- **Corner markers, not full borders** — L-brackets are the genre signature. Full border looks like a generic UI box.
- **Confidence label flickers in a tight range (CONFIDENCE_MEAN ± CONFIDENCE_VAR inside [95, 99])** — outside that range reads as "uncertain"; ≥100 reads as "fake-precise."
- **No CSS animation for the tracking — use timeline onUpdate** — HF seek-by-frame doesn't sync with CSS animation.
## Critical Constraints
- **Timeline must be paused**: `gsap.timeline({ paused: true })`
- **Registry key = `data-composition-id`**
- **No CSS animation on `.track-box` or `.corner`** — must be timeline-driven
- **`will-change: transform, width, height`** on `.track-box`
- **`pointer-events: none`** on `.track-box` — decorative overlay
- **Box position recomputed per-frame from target** — never tween box position separately from target
## Combinations
- [viewport-change.md](viewport-change.md) — zoom into the tracked box after detection phase
- [multi-phase-camera.md](multi-phase-camera.md) — wide shot during tracking, push-in on lock
- [sine-wave-loop.md](sine-wave-loop.md) — the mascot itself idle-breathes inside the box
## Pairs with HF skills
- `/hyperframes-animation` — onUpdate writing multi-element positions
- `/hyperframes-core` — composition wiring
- `/hyperframes-cli` — `hyperframes lint`