CLAIM: Camera-LiDAR Alignment with Intensity and Monodepth
This work addresses the limitations of manual feature matching and poor robustness in camera–LiDAR extrinsic calibration. We propose an end-to-end, feature-free coarse-to-fine alignment method. Our approach innovatively fuses LiDAR intensity images with monocular depth predictions to establish a dual-loss alignment framework: a structural loss based on patch-wise Pearson correlation ensures geometric consistency, while a mutual information-based texture loss enhances radiometric consistency. To incorporate strong geometric priors, we leverage a pre-trained monocular depth model and design a lightweight spatial search optimization framework. The method exhibits scene adaptability and cross-dataset generalizability. Extensive experiments on KITTI, Waymo, and MIAS-LCEC demonstrate significant improvements over state-of-the-art methods in both calibration accuracy and robustness. Code is publicly available.