We propose a novel data augmentation method for point cloud, Rigid Subset Mix (RSMix). Our model is implemented based on PointNet+++ and DGCNN, which are widely used point-wise deep neural networks.

• [CVPR paper Link]

• [CVPR supplementary Link]

• [arXiv version paper link]

RSMix generates the virtual sample from each part of the two point cloud samples by mixing them without shape distortion. It effectively generalize the deep neural network model and achieve remarkable performance for shape classification.

Qualitative samples of RSMix.

Implementation of RSMix on PointNet++ implemented in TensorFlow.

• RSMix-PointNet++(TensorFlow)

Implementation of RSMix on DGCNN implemented in PyTorch.

• RSMix-DGCNN(PyTorch)

Please refer the following repos. Both repo include same visualizer proposed by [PointNet++]. In addition, second repo also include visualization tool using [open3d]. For the pure visualization for point cloud, please refer the pure visualization code on second repo.

• RSMix-PointNet++(TensorFlow)

• Point Cloud Visualizer

ModelNe40-C, a new corruption robustness benchmark for data augmentation in point cloud is recently proposed. Our work is also tested on the dataset. They found that RSMix is robust to "density" corruptions. If you are interested in the results, please refer the follow link.

• [ModelNet40-C github]
• [ModelNet40-C paper]

MIT License

The structure of this codebase is borrowed from PointNet++ and DGCNN-PyTorch.

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

@inproceedings{lee2021regularization,
  title={Regularization strategy for point cloud via rigidly mixed sample},
  author={Lee, Dogyoon and Lee, Jaeha and Lee, Junhyeop and Lee, Hyeongmin and Lee, Minhyeok and Woo, Sungmin and Lee, Sangyoun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={15900--15909},
  year={2021}
}