Train lidar_apollo_instance_segmentation CNN

Train lidar_apollo_instance_segmentation CNN with Nuscenes.

Getting Started

Only supports python3.

1. Clone original_model branch and install some packages.

git clone --recursive https://github.com/kosuke55/train_baiducnn.git
pip install -r requirements.txt
mkdir build
cd build
cmake ..
make -j

2. create_dataset_from_nusc.py is for creating a dataset to train apollo cnn. Set dataroot and save_dir.

cd ~/train_baiducnn/scripts/create_dataset  
python create_dataset_from_nusc.py  --dataroot <downloaded nuscenes path> --save_dir <dir to save created dataset> --nusc_version <v1.0-mini or v1.0-trainval>  

3. Execute start_server.sh and access from a web browser. Then you can train with train_bcnn.py.

cd ~/train_baiducnn/scripts/pytorch  
./start_server.sh  
python train_bcnn.py --data_path <dir to save created dataset>  

4. Trained model can be converted to onnx by pytorch2onnx.py and converted to engine by onnx-tensorrt.

cd ~/train_baiducnn/scripts/pytorch  
python pytorch2onnx --trained_model <your_trained_model.pt>  
# after installing onnx-tensorrt  
onnx2trt <your_trained_model.onnx> -o <your_trained_model.engine>  

5. Run lidar_apollo_instance_segmentation with <your_trained_model.engine>

reference

apollo 3D Obstacle Percption description

autoware_perception description

yunlongdong/FCN-pytorch onehot