RCP: Recurrent Closest Point for Scene Flow Estimation on 3D Point Clouds

This is the official PyTorch implementation code for RCP. For technical details, please refer to:

RCP: Recurrent Closest Point for Scene Flow Estimation on 3D Point Clouds
Xiaodong Gu, Chengzhou Tang, Weihao Yuan, Zuozhuo Dai, Siyu Zhu, Ping Tan
[Paper]

Installation

Install python dependencies lib:

pip install -r requirements.txt

Install PointNet2 CPP lib:

cd lib/pointnet2
python3 setup.py install

Datasets

We follow HPLFlowNet preprocessing methods:

FlyingThings3D: Download and unzip the "Disparity", "Disparity Occlusions", "Disparity change", "Optical flow", "Flow Occlusions" for DispNet/FlowNet2.0 dataset subsets from the FlyingThings3D website (we used the paths from this file, now they added torrent downloads) . They will be unzipped into the same directory, RAW_DATA_PATH. Then run the following script for 3D reconstruction:

python data/preprocess/process_flyingthings3d_subset.py --raw_data_path ${RAW_DATA_PATH} --save_path ${SAVE_PATH}/FlyingThings3D_subset_processed_35m --only_save_near_pts

KITTI: Download and unzip KITTI Scene Flow Evaluation 2015 to directory RAW_DATA_PATH. Run the following script for 3D reconstruction:

python data/preprocess/process_kitti.py ${RAW_DATA_PATH} ${SAVE_PATH}/KITTI_processed_occ_final

Training

Fully-supervised training:

python run.py -c configs/train/rcp_sup_pre.yaml
python run.py -c configs/train/rcp_sup_ft.yaml --pre_ckpt ${pretrained_ckpt}

Self-supervised training:

python run.py -c configs/train/rcp_self_pre.yaml
python run.py -c configs/train/rcp_self_ft.yaml --pre_ckpt ${pretrained_ckpt}

Evaluation

Evaluate on FlyingThings3D

python run.py -c configs/test/rcp_test.yaml --test_ckpt ${ft_ckpt}

Evaluate on KITTI

python run.py -c configs/test/rcp_test_kitti.yaml --test_ckpt ${ft_ckpt}

Pretrained Models

Download Link

Datasets	EPE3D	Acc3DS	AccDR	Outliers3D
FlyingThings3D	0.0403	0.8567	0.9635	0.1976
KITTI	0.0481	0.8491	0.9448	0.1228

Citation

If you find this code useful in your research, please cite:

@inproceedings{gu2022rcp,
  title={RCP: Recurrent Closest Point for Point Cloud},
  author={Gu, Xiaodong and Tang, Chengzhou and Yuan, Weihao and Dai, Zuozhuo and Zhu, Siyu and Tan, Ping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8216--8226},
  year={2022}
}

Acknowledgements

Some code are borrowed from Flowstep3d, FLOT, flownet3d_Pytorch, HPLFlowNet and Pointnet2.PyTorch. Thanks for these great projects.

AlibabaResearch/rcp