Simple-Calib: An End-to-End Network for Simultaneous Calibration of Multi LiDAR-Camera

Multi-camera and multi-LiDAR configurations are widely adopted in autonomous driving to improve environmental perception. However, existing LiDAR-Camera calibration networks are typically limited to one-to-one sensor pairs, resulting in poor scalability and high computational costs. To address these limitations, we propose a novel calibration network capable of simultaneously calibrating multi-camera and multi-LiDAR in a single forward pass. Our approach employs pre-trained Swin-Transformers as shared backbones to extract features from both camera images and LiDAR-generated depth maps. Two Feature Pyramid Networks are used to enhance multiscale representations, and A dedicated calibration head concurrently predicts the extrinsics for six LiDAR-Camera pairs. Experimental results on the nuScenes dataset show that our model, with only 75.6 M parameters, achieves a mean translation error of 3.429 cm and a rotation error of 0.509°. While slightly less accurate in calibration precision, it significantly outperforms existing approaches in terms of scalability and model compactness.