Transformed reprojection error: a sample selection metric for camera - LiDAR calibration for driving scenes

Official URL: https://dx.doi.org/10.1109/ICCE67443.2026.11449699

In this study, we introduce a novel sample selection metric, the Transformed Reprojection Error (TRE), designed to improve the accuracy of traditional camera-lidar calibration pipelines. Using a dataset acquired from a controlled calibration room, we perform multiple iterations of the camera-lidar calibration loop, varying the number of frames used in each iteration to evaluate calibration quality based on the newly introduced metric. The final calibration parameters are selected based on the minimum transformed reprojection error observed across the iterations. To validate our approach, we test the calibrated parameters on a vehicle placed at a far range in a real driving scene. The results demonstrate that our proposed metric significantly enhances calibration accuracy compared to traditional methods, providing improved sensor alignment in both controlled environments and verified real-world scenarios. This makes the TRE a valuable addition to camera-lidar calibration workflows, offering higher alignment of LiDAR points onto the images through the iterative optimization process.