perhapswo/centerpointtensorrt

A trt version of CenterPoint model for Lidar Detection

An Lidar Object Detection project implemented by TensorRT

The project is a TensorRT version of CenterPoint, an 3D object detection model using center points in the bird eye view. Code is written according to the project

Besides, it is running inference on WaymoOpenSet

Setup

The project has been tested on Ubuntu18.04 and Ubuntu20.04, It mainly relies on TensorRT and cuda as 3rd-party package, with the following versions respectively:

vTensorRT : 8.0.1.6

vCuda : 11.3

Note that this project does not rely on PCL and Boost by now. However, they may be used in the future and has been written in CMakeLists.txt.

After installation, you may then build the project by executing the following commands:

cd /YOUR/PATH/TO/centerpoint
mkdir centerpoint_pp_baseline_score0.1_nms0.7 && cd src
mkdir build && cd build
cmake .. && make
./centerpoint

By default, the project loads the serialized engine files to do inference, and the engine files are created by the onnx files we provided and are set as float16. You can also build from onnx files by setting params.load_engine = false in samplecenterpoint.cpp and provide the onnx file path. In that way, you may decide whether to use fp16 or fp32.

What has been done?

To futher learn the detailed documentation, please refer to the following computation graph and doc file.

Computation Speed

Acceleration is the main aim we want to archieve, and therefore we do most of computation(including preprocess & postprocess) on GPU. The below table gives the average computation speed (by millisecond) of every computation module, and it is tested on RTX3080, with all the 39987 waymo validation samples. As illustrated above, the engine is set to float16, and eveluation metric shows no difference on fp32 or fp16.

Preprocess	PfeInfer	ScatterInfer	RpnInfer	Postprocess
1.61	5.88	0.17	6.89	2.37

Detection result shows below:

Acknowledgements

This project refers to some codes from:

CenterPoint

TensorRT

CenterPoint-PointPillars