Unsupervised Learning of 3D Scene Flow with LiDAR Odometry Assistance

Environment setup

Tested with Pytorch 1.10.0 and CUDA 11.3

1. conda create -n psf python=3.8 -y
   conda activate psf
   pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio==0.10.2 -f https://download.pytorch.org/whl/torch_stable.html
   pip install setuptools==58.2.0
   pip install pandas addict tqdm scipy pytz

2. Install pointnet2

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

   cd pointnet2
   python setup.py install

Train

Set dataset_path in the configuration file to the path of the dataset. Then run the following command to train the model.

python train.py --config config/cfg_psf_train.py

Evaluation

Set resume_from in the configuration file to the path of the dataset. Then run the following command to evaluate the model.

python eval.py --config config/cfg_psf_eval.py

Result

Visualization of scene flow estimation results on the K-SF dataset for our method and baseline method PointPWC-Net.

Citation

If you find our work useful in your research, please cite:

@article{wang2022unsupervised,
  title={Unsupervised learning of 3D scene flow with 3D odometry assistance},
  author={Wang, Guangming and Feng, Zhiheng and Jiang, Chaokang and Wang, Hesheng},
  journal={arXiv preprint arXiv:2209.04945},
  year={2022}
}

Acknowledgements

We thank the following open-source projects for the help of the implementations.

• PointNet++

• PointPWC-Net

We also provided the links to the baseline methods used in our paper as follows:

• ICP
• FGR
• CPD
• Ego-motion
• PointPWC-Net
• JGwF
• SFGAN
• FLOT
• Jin et al.
• SPFlowNet