Contrastive Boundary Learning for Point Cloud Segmentation (CVPR 2022)

By Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao

This is the implementation of our CVPR 2022 paper:
Contrastive Boundary Learning for Point Cloud Segmentation [arXiv]

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

@InProceedings{tang2022cbl,
    author    = {Tang, Liyao and Zhan, Yibing and Chen, Zhe and Yu, Baosheng and Tao, Dacheng},
    title     = {Contrastive Boundary Learning for Point Cloud Segmentation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {8489-8499}
}

Setup & Usage

For point-transformer baseline, please follow pytorch/README.

For ConvNet and other baselines, please follow tensorflow/README.

Pre-trained models

Pretrained models can be accessed here, together with training and testing log. Choose the desired baseline and unzip into the corresponding code directory (tensorflow/pytorch) and follow the README there for further instruction.

Quantitative results

S3DIS (Area 5)

baseline	mIoU	OA	mACC
ConvNet + CBL	69.4	90.6	75.2
ConvNet + CBL (kl)	69.5	90.9	75.3
point-transformer + CBL	71.6	91.2	77.9

Qualitative results

Acknowledgement

Codes are built based on a series of previous works, including:
KPConv,
RandLA-Net,
CloserLook3D,
Point-Transformer.
Thanks for their excellent work.

License

This repo is licensed under the terms of the MIT license (see LICENSE file for details).