Awesome

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 <br /> Xiaodong Gu, Chengzhou Tang, Weihao Yuan, Zuozhuo Dai, Siyu Zhu, Ping Tan <br /> [Paper] <br />

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

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.