M3Net: Multi-Space Alignments Towards Universal LiDAR Segmentation

Youquan Liu^*,1    Lingdong Kong^*,2,3    Xiaoyang Wu⁴    Runnan Chen⁴    Xin Li⁵    Liang Pan²    Ziwei Liu⁶    Yuexin Ma^1   
1ShanghaiTech University    ²Shanghai AI Laboratory    ³National University of Singapore    ⁴University of Hong Kong    ⁵East China Normal University    ⁶S-Lab, Nanyang Technological University   

About

M3Net is a new type of LiDAR segmentation network that unifies the multi-task, multi-dataset, and multi-modality learning objectives.

News

• [2024.05] - Our paper is available on arXiv, click here to check it out.
• 🔥[2024.02] - M3Net was accepted to CVPR 2024!

Getting Started

Please refer to GET_STARTED.md to learn more about how to use this codebase.

Qualitative Evaluation

SemanticKITTI

nuScenes

Waymo Open

Citation

If you find this work helpful, please kindly consider citing our paper:

@inproceedings{liu2024multi,
  title={Multi-Space Alignments Towards Universal LiDAR Segmentation},
  author={Liu, Youquan and Kong, Lingdong and Wu, Xiaoyang and Chen, Runnan and Li, Xin and Pan, Liang and Liu, Ziwei and Ma, Yuexin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14648--14661},
  year={2024}
}

Acknowledgements

The overall structure of this repo is derived from Pointcept, SAM, OpenSeed and OpenPCSeg. Thank the authors for their great work!