To build the transfuison inference, TensorRT and CUDA are needed.
- Ubuntu 22.04
- CUDA 12.1 + cuDNN 8.9.2.26-1+cuda12.1 + TensorRT 8.6.1.6-1+cuda12.0
$ sudo apt install libopen3d-dev
$ git clone https://github.com/wep21/CUDA-TransFusion.git && cd CUDA-TransFusion
$ cmake -Bbuild
$ cmake --build build -j$(nproc)
$ cmake --install build --prefix install
$ /usr/src/tensorrt/bin/trtexec --onnx=sample/transfusion.onnx --minShapes=voxels:5000x20x5,num_points:5000,coors:5000x4 --optShapes=voxels:10000x20x5,num_points:10000,coors:10000x4 --maxShapes=voxels:30000x20x5,num_points:30000,coors:30000x4 --saveEngine=sample/transfusion.plan
$ export LD_LIBRARY_PATH=install/lib:$LD_LIBRARY_PATH
$ install/bin/transfusion_main sample/n015-2018-07-24-11-22-45+0800__LIDAR_TOP__1532402927647951.pcd.bin sample/transfusion.plan