C-LaV: Conditional Latent Velocity Field Denoising for Weather-Robust LiDAR Place Recognition
Official model weights for C-LaV (CVPR 2026).
- 📄 Project page: https://patience-joey.github.io/clav/
- 💻 Code: https://github.com/Patience-Joey/clav
- 🎓 Authors: Xuewei Cao, Jiayue Yang, Zhiwen Zeng, Yanyong Zhang, Yan Xia (University of Science and Technology of China)
Quick start
git clone https://github.com/Patience-Joey/clav.git
cd clav
conda env create -f environment.yml && conda activate clav
# Download weights from this repo
pip install huggingface_hub
python -c "
from huggingface_hub import hf_hub_download
for d in ('kitti', 'nclt', 'boreas'):
for f in ('stage2.pt', 'best.pt'):
p = hf_hub_download('xueweicao/clav', f'{d}/{f}')
print(p)
"
# Evaluate
bash scripts/eval/evaluate_kitti.sh --checkpoint <kitti/best.pt>
bash scripts/eval/evaluate_nclt.sh --checkpoint <nclt/best.pt>
bash scripts/eval/evaluate_boreas.sh --checkpoint <boreas/best.pt>
Datasets
Trained / evaluated on:
- KITTI — synthetic rain/fog/snow on the original clear-weather scans (Hahner et al. fog/snow simulation)
- NCLT — cross-session evaluation with synthetic adverse weather
- Boreas — real-world rain and snow, cross-pass GPS-aligned pairs
License & citation
Released under MIT. If C-LaV helps your work, please cite:
@inproceedings{cao2026clav,
title = {C-LaV: Conditional Latent Velocity Field Denoising for Weather-Robust LiDAR Place Recognition},
author = {Cao, Xuewei and Yang, Jiayue and Zeng, Zhiwen and Zhang, Yanyong and Xia, Yan},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2026}
}