Table Of Contents
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.
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.
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. |
The following resources provide a deeper understanding of the pillarScatterPlugin plugin:
Networks:
Documentation:
For terms and conditions for use, reproduction, and distribution, see the TensorRT Software License Agreement documentation.
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.
There are no known issues in this plugin.