CenterPoint inferenced with TensorRT

This repository contains sources and model for CenterPoint inference using TensorRT. The model is created with mmdetection3d.

Overall inference has five phases:

• Convert points cloud into 4-channle voxels
• Extend 4-channel voxels to 10-channel voxel features
• Run pfe TensorRT engine to get 64-channel voxel features
• Run rpn backbone TensorRT engine to get 3D-detection raw data
• Parse bounding box, class type and direction

Model && Data

The demo used the custom dataset like KITTI. The onnx file can be converted by onnx_tools

Prerequisites

To build the centerpoint inference, TensorRT and CUDA are needed.

Environments

• NVIDIA RTX A4000 Laptop GPU
• CUDA 11.1 + cuDNN 8.2.1 + TensorRT 8.2.3

Compile && Run

$ mkdir build && cd build
$ cmake .. && make -j$(nproc)
$ ./demo

Visualization

You should install open3d in python environment.

$ cd tools
$ python viewer.py

trt fp16	pytorch

Performance in FP16

| Function(unit:ms) | NVIDIA RTX A4000 Laptop GPU | NVIDIA Jetson AGX Orin      |
| ----------------- | --------------------------- | --------------------------- |
| Preprocess        | 0.950476 ms                 | 3.52855  ms                 |
| Pfe               | 4.37507  ms                 | 18.2881  ms                 |
| Scatter           | 0.204093 ms                 | 1.33041  ms                 |
| Backbone          | 9.84435  ms                 | 20.7511  ms                 |
| Postprocess       | 2.91952  ms                 | 4.61471  ms                 |
| Summary           | 18.2961  ms                 | 48.5218  ms                 |

Note

• The pretrained model in this project doesn't predict vx, vy.
• The demo will cache the onnx file to improve performance. If a new onnx will be used, please remove the cache file in "./model".

References

• Center-based 3D Object Detection and Tracking
• mmdetection3d
• tianweiy/CenterPoint
• Abraham423/CenterPoint