yuanwenyue/LitePT

This repository contains model weights for LitePT: Lighter Yet Stronger Point Transformer, a lightweight, high-performance 3D point cloud architecture. LitePT proposes an improved 3D point cloud backbone that employs convolutions in early stages and switches to attention for deeper layers, introducing a novel, training-free 3D positional encoding, PointROPE. The resulting model achieves state-of-the-art performance while being significantly more efficient.

Paper & Resources

• Paper: LitePT: Lighter Yet Stronger Point Transformer
• Arxiv: https://arxiv.org/abs/2512.13689
• Project Page: https://litept.github.io/
• Codebase: https://github.com/prs-eth/LitePT

Models

We release the pretrained model weights for the benchmarks we reported in our paper.

Semantic segmentation

Model	Params	Benchmark	val mIoU	Config	Checkpoint
LitePT-S	12.7M	NuScenes	82.2	link	Download
LitePT-S	12.7M	Waymo	73.1	link	Download
LitePT-S	12.7M	ScanNet	76.5	link	Download
LitePT-S	16.0M	Structured3D	83.6	link	Download
LitePT-B	45.1M	Structured3D	85.1	link	Download
LitePT-L	85.9M	Structured3D	85.4	link	Download

Instance segmentation

Model	Params	Benchmark	mAP25	mAP50	mAP	Config	Checkpoint
LitePT-S*	16.0M	ScanNet	78.5	64.9	41.7	link	Download
LitePT-S*	16.0M	ScanNet200	40.3	33.1	22.2	link	Download

Object detection

Model	Params	Benchmark	mAPH	Config	Checkpoint
LitePT	9.0M	Waymo	70.7	link	Download

Citation

@article{yuelitept2025,
    title={{LitePT: Lighter Yet Stronger Point Transformer}},
    author={Yue, Yuanwen and Robert, Damien and Wang, Jianyuan and Hong, Sunghwan and Wegner, Jan Dirk and Rupprecht, Christian and Schindler, Konrad},
    journal={arXiv preprint arXiv:2512.13689},
    year={2025}
}