English | 简体中文

PointPillars

High performance version of 3D object detection network -PointPillars, which can achieve the real-time processing (less than 1 ms / head)

The inference part of PointPillars(pfe , backbone(multihead)) is optimized by tensorrt
The pre- and post- processing are optimized by CUDA / C + recode.

Major Advance

Easy to train
- this repo directly uses mmlab/mmdetection3d for training. Therefore, as long as you follow the steps of the official tutorial, it is very easy to train your own data. In addition, you can directly use official weight(hv_pointpillars_secfpn_sbn-all_4x8_2x_nus-3d) for deployment. Due to the need to use Tensorrt, the official mmlab/mmdetection3d still needs to be customized. You can find onnx export tool in speshowBUAA/mmdet3d_onnx_tools.
Easy to deploy
- this repo is mainly modified to be compatible with mmdet3d models on the basis of hova88/PointPillars_MultiHead_40FPS

clone thest two repositories, and make sure the dependences is complete

mkdir workspace && cd workspace
git clone https://github.com/speshowBUAA/PointPillars_mmdet_secfpn.git --recursive && cd ..

generate engine file
- 1.1 Pytorch model --> ONNX model : Please refer to speshowBUAA/mmdet3d_onnx_tools.
- 1.2 ONNX model --> TensorRT model : after install the onnx2trt, things become very simple. Note that if you want to further improve the the inference speed, you must use half precision or mixed precision(like ,-d 16)
```
    onnx2trt pts_pfe.onnx -o pts_pfe.trt -b 1 -d 16 
    onnx2trt pts_backbone.onnx -o pts_backbone.trt -b 1 -d 16 
```
- 1.3 engine file --> algorithm : Specified the path of engine files(*.onnx , *.trt) inbootstrap.yaml.
- 1.4 Download the test pointcloud nuscenes_10sweeps_points.txt, and specified the path in bootstrap.yaml.

Compiler

cd PointPillars_mmdet_secfpn
mkdir build && cd build
cmake .. && make -j8 && ./test/test_model

Visualization

cd PointPillars_mmdet_secfpn/tools
python viewer.py

GNU General Public License v3.0 or later See COPYING to see the full text.