This repository is the official implementation of the paper:
[CVPR 2025] CamPoint: Boosting Point Cloud Segmentation with Virtual Camera.
TL;DR: The CamPoint employs Camera Visibility Feature(CVF) to encode points as feature vector via virtual cameras, representing the visibility from multiple camera views. Mainly works include:
- Camera Perspective Slice Distance(CPSD): Identifies semantically related neighbors rather than just spatially closest points to enhance local feature aggregation.
- Camera Parameter Embedding(CPE): Integrates camera prior features into point representations to enhance global information perception.
The CamPoint achieves SOTA performance on multiple datasets (e.g., 83.3% mIoU on S3DIS, without any spacial strategies like voting, pre-training, or joint training), with fewer parameters, lower training costs, and faster inference speed.
Setup python environment: SETUP.md
Train and test datasets:
- checkpoints and train logs: https://huggingface.co/MTXAI/CamPoint/tree/main/exp
- test logs: https://huggingface.co/MTXAI/CamPoint/tree/main/exp-test
- tensorboard: https://huggingface.co/MTXAI/CamPoint/tensorboard
If you find CamPoint method or codebase useful, please cite:
- PointNext: https://github.com/guochengqian/PointNeXt
- mamba: https://github.com/state-spaces/mamba
- DeLA: https://github.com/Matrix-ASC/DeLA
- msplat: https://github.com/pointrix-project/msplat
- gaussian-splatting: https://github.com/graphdeco-inria/gaussian-splatting
- pytorch3d: https://pytorch3d.org/
- kiui: https://kit.kiui.moe/