OpenPCDet trains PointPillar in KITTI

Table of contents

1. Environment: pytorch1.2

2. Download the source code:

3. Prepare the KITTI dataset

1. Download these (very huge)

 2. After downloading, arrange according to the following directory

Fourth, run the code

1. Perform a wave of data processing:

 2. start running

2.1 Error: KeyError: 'road_plane'

2.2报错：LoweringError: Failed in nopython mode pipeline (step: nopython mode backend)

3. Complete

---

1. Environment: pytorch1.2

2. Download the source code:

#下载源码
git clone https://github.com/open-mmlab/OpenPCDet.git

#环境配置
cd OpenPCDet
python setup.py develop
pip install -r requirements.txt

3. Prepare the KITTI dataset

2017 website: The KITTI Vision Benchmark Suite

1. Download these (very huge)

 2. After downloading, arrange according to the following directory

Fourth, run the code

1. Perform a wave of data processing:

python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/dataset_configs/kitti_dataset.yaml

 2. start running

cd tools
python train.py --cfg_file=cfgs/kitti_models/pointpillar.yaml --batch_size=3 --epochs=100

2.1 Error: KeyError: 'road_plane'

Solution: For the file ./tools/cfgs/kitti_models/pointpillar.yaml

Change line 27 USE_ROAD_PLANE: True to USE_ROAD_PLANE: False

It's ok to run again

2.2报错：LoweringError: Failed in nopython mode pipeline (step: nopython mode backend)

Solution: The numba version is wrong, replace it with 0.51.0

3. Complete

The result is stored in

./output/kitti_models/pointpillar/default/eval/eval_with_train/epoch_100/val/result.pkl

./output/kitti_models/pointpillar/default/ckpt/checkpoint_epoch_100.pth