Snow-Calib: Deep Learning-Based Camera-LiDAR Extrinsic Calibration Under Snowy Weather Conditions
Lists
Only you see this
Abstract
The fusion of camera and LiDAR sensors necessitates precise extrinsic calibration; however, existing deep learning methods are often computationally inefficient and perform poorly in dynamic and adverse environments, such as under snowfall. To address these limitations, we propose a novel framework called Snow-Calib that dynamically integrates multimodal data via an optimized architecture, embedding image-based snow removal via attention mechanisms and multi-scale feature fusion to suppress weather-induced noise while preserving structural details. By jointly leveraging the LiDAR’s geometric robustness and denoised RGB imagery, the framework enhances the accuracy of cross-modal feature matching without requiring separate pre-processing. The overall proposed network consists of 11.9 million parameters, categorized as a lightweight calibration network that can reduce computational overhead. Therefore, the model is resilient to environmental distortions and consumes fewer resources, critical for deployment in autonomous systems operating under variable real-world conditions. Experimental results reveal that Snow-Calib achieves an average calibration error of 0.856 cm for translation and 0.23° for rotation. The code for Snow-Calib is publicly available at https://github.com/MFMummtaz/Snow-Calib.
BibTeX
@inproceedings{mummtaz2025snow,
author = {Muhammad Fairuz Mummtaz and Jaejun Yoo and Muhammad Rangga Aziz Nasution and Ida Bagus Krishna Yoga Utama and Miftahul Khoir Shilahul Umam and Muhammad Alfi Aldolio and Su Mon Ko and Yeong Min Jang},
title = {{Snow-Calib: Deep Learning-Based Camera-LiDAR Extrinsic Calibration Under Snowy Weather Conditions}},
booktitle = {IEEE Access},
year = {2025},
doi = {10.1109/access.2025.3624453},
}Loading related content...