Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation

Paper (arXiv 2505.23926) · Project Page · Code · Google Drive mirror

Pretrained Point-MoE-L checkpoints for the paper Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation (ICLR 2026).

Checkpoints

Checkpoint	Config	Training data
pointmoe_l_indoor_last.pth	pointmoe_l_indoor_last.py	ScanNet, Structured3D, S3DIS (Area 1/2/3/4/6)
pointmoe_l_indoor_outdoor_last.pth	pointmoe_l_indoor_outdoor_last.py	ScanNet, Structured3D, S3DIS (Area 1/2/3/4/6), SemanticKITTI, nuScenes

All models are trained on the official training splits only.

Model: Point Transformer V3 backbone with Mixture-of-Experts layers (8 experts, top-2 routing). ~100M total parameters, ~59M activated per token.

Usage

1. Set up the environment and prepare the datasets under data/ following the code repository.
2. Download a checkpoint and its config:

huggingface-cli download uva-cv-lab/Point_MoE pointmoe_l_indoor_last.pth pointmoe_l_indoor_last.py --local-dir checkpoints

3. Evaluate from the root of the code repository:

python tools/train.py \
  --config-file checkpoints/pointmoe_l_indoor_last.py \
  --num-gpus 4 \
  --options weight=checkpoints/pointmoe_l_indoor_last.pth save_path=exp/pointmoe_l_indoor

Citation

@inproceedings{chenpoint,
  title={Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation},
  author={Chen, Xuweiyi and Zhou, Wentao and RoyChowdhury, Aruni and Cheng, Zezhou},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

Acknowledgements

Built on Pointcept.