wehub-resource-sync
86 lines
3.4 KiB
Markdown
86 lines
3.4 KiB
Markdown
# pillarScatter Plugin [DEPRECATED]
|
|
|
|
**This plugin is deprecated since TensorRT 10.12 and will be removed in a future release. No alternatives are planned to be provided.**
|
|
|
|
**Table Of Contents**
|
|
- [Description](#description)
|
|
* [Structure](#structure)
|
|
- [Parameters](#parameters)
|
|
- [Additional resources](#additional-resources)
|
|
- [License](#license)
|
|
- [Changelog](#changelog)
|
|
- [Known issues](#known-issues)
|
|
|
|
## Description
|
|
|
|
The `pillarScatterPlugin` performs scatter of voxels for PointPillars model. This operation is roughly a sparse to dense conversion and similar to ordinary scatter operation. The difference is `pillarScatterPlugin` will only scatter voxels that are marked as valid and do nothing for invalid voxels.
|
|
|
|
`pillarScatterPlugin` implements a sparse to dense conversion of voxels. The plugin takes sparse voxel values and their corresponding indices and produces a dense representation of the voxels(pillars).
|
|
|
|
This plugin is optimized for the above steps and it allows you to do PointPillars inference in TensorRT.
|
|
|
|
|
|
### Structure
|
|
|
|
The `pillarScatterPlugin` takes 3 inputs; `voxels`, `voxel_coords`, and `num_pillar`.
|
|
|
|
`voxels`
|
|
The tensor of voxel values. The shape of this tensor is `[N, P, C]`, where `N` is the batch size, `P` is the maximum number of pillars per frame, and `C` is the number of features for each pillar.
|
|
|
|
|
|
`voxel_coords`
|
|
The dense coordinates of `voxels`. This tensor contains the coordinate of each voxel in `voxels`. The coordinates will determine which voxel goes to which index in the output dense tensor. This tensor has a shape of `[N, P, 4]`, where `N, P` are as above and `4` is the length of the coordinates encoded as `(frame_id, z, y, x)`.
|
|
|
|
|
|
`num_pillar`
|
|
The number of valid pillars. For each frame, the number of voxels(pillars) can be lower than the maximum pillar number `P`. Only valid pillars will be considered in this operation. Invalid pillars will simply be ignored. The dimension of this tensor is simply `[N]`.
|
|
|
|
|
|
The `pillarScatterPlugin` generates the following 1 output:
|
|
|
|
`dense_feature_map`
|
|
The result of this operation is a dense feature map with the shape `[N, C, H, W]`, where `N, C` are as above and `H, W` are the height and width of the dense feature map. This tensor usually is a highly sparse tensor.
|
|
|
|
|
|
## Parameters
|
|
|
|
`pillarScatterPlugin` has plugin creator class `pillarScatterPluginCreator` and plugin class `pillarScatterPlugin`.
|
|
|
|
The parameters are defined below and consists of the following attributes:
|
|
|
|
| Type | Parameter | Description
|
|
|----------|--------------------------|--------------------------------------------------------
|
|
|`list of ints` | `dense_shape` | The shape(h, w) of the dense feature map that will be generated by this operation.
|
|
|
|
## Additional resources
|
|
|
|
The following resources provide a deeper understanding of the `pillarScatterPlugin` plugin:
|
|
|
|
**Networks:**
|
|
- [PointPillars](https://arxiv.org/pdf/1812.05784)
|
|
|
|
**Documentation:**
|
|
- [PointPillars](https://arxiv.org/pdf/1812.05784)
|
|
|
|
## License
|
|
|
|
For terms and conditions for use, reproduction, and distribution, see the [TensorRT Software License Agreement](https://docs.nvidia.com/deeplearning/sdk/tensorrt-sla/index.html)
|
|
documentation.
|
|
|
|
|
|
## Changelog
|
|
|
|
May 2025
|
|
Add deprecation note.
|
|
|
|
Feb 2023
|
|
Fixed a bug where PillarScatter would incorrectly report support for `HWC8` formats.
|
|
|
|
Dec 2021
|
|
This is the first release of this `README.md` file.
|
|
|
|
|
|
## Known issues
|
|
|
|
There are no known issues in this plugin.
|