Skip to content

PointPWC-Net is a deep coarse-to-fine network designed for 3D scene flow estimation from 3D point clouds.

License

Notifications You must be signed in to change notification settings

CyrilSterling/PointPWC

 
 

Repository files navigation

PointPWC

This is the code for PointPWC-Net, a deep coarse-to-fine network designed for 3D scene flow estimation from 3D point clouds.

Prerequisities

Our model is trained and tested under:

  • Python 3.6.9
  • NVIDIA GPU + CUDA CuDNN
  • PyTorch (torch == 1.5)
  • scipy
  • tqdm
  • sklearn
  • numba
  • cffi
  • pypng
  • pptk

Compile the furthest point sampling, grouping and gathering operation for PyTorch. We use the operation from this repo.

cd pointnet2
python setup.py install
cd ../

Data preprocess

For fair comparison with previous methods, we adopt the preprocessing steps in HPLFlowNet. Please refer to repo. We also copy the preprocessing instructions here for your reference.

  • 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 upzipped into the same directory, RAW_DATA_PATH. Then run the following script for 3D reconstruction:
python3 data_preprocess/process_flyingthings3d_subset.py --raw_data_path RAW_DATA_PATH --save_path SAVE_PATH/FlyingThings3D_subset_processed_35m --only_save_near_pts
python3 data_preprocess/process_kitti.py RAW_DATA_PATH SAVE_PATH/KITTI_processed_occ_final

Get started

Here are some demo results:

Train

Set data_root in the configuration file to SAVE_PATH in the data preprocess section. Then run

python3 train.py config_train.yaml

Evaluate

Set data_root in the configuration file to SAVE_PATH in the data preprocess section. Then run

python3 evaluate.py config_evaluate.yaml

We upload one pretrained model in pretrain_weights.

Citation

If you use this code for your research, please cite our paper.

@article{wu2019pointpwc,
  title={PointPWC-Net: A Coarse-to-Fine Network for Supervised and Self-Supervised Scene Flow Estimation on 3D Point Clouds},
  author={Wu, Wenxuan and Wang, Zhiyuan and Li, Zhuwen and Liu, Wei and Fuxin, Li},
  journal={arXiv preprint arXiv:1911.12408},
  year={2019}
}

Acknowledgement

We thank repo and repo for subsampling, grouping and data preprocessing related functions.

About

PointPWC-Net is a deep coarse-to-fine network designed for 3D scene flow estimation from 3D point clouds.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 78.5%
  • Cuda 13.3%
  • C++ 8.2%