Weakly Supervised Geometric Self-Consistency for LiDAR-Camera Calibration

LiDAR-Camera data fusion relies heavily on high-precision calibration of extrinsic parameters. Most existing learning-based approaches heavily depend on numerous manually annotated labels, which pose significant challenges for practical deployment, particularly for rotation angles. In this paper, we propose a weakly supervised LiDAR-Camera calibration framework to implement joint calibration between Camera and LiDAR, named GCCalib, which leverages geometric self-consistency between various input modalities to guide the model's self-supervised rotation estimation. Furthermore, to mitigate quantization error during projection, we design a spatial quantization compensation module. Additionally, we introduce a self-attention mechanism for correlation volume computation to enhance the feature extraction capabilities. Comprehensive experiments on benchmark datasets demonstrate that our approach achieves state-of-the-art performance in weakly supervised methods, even outperforms some supervised methods. Our approach reduces dependence on ground truth data and enhances model generalization, thereby facilitating real-time online calibration.