Learning Which Correspondences to Trust: Confidence-Weighted Event-Camera Localization in LiDAR Maps
The method CELL improves event camera localization against LiDAR maps by learning per-correspondence confidence end-to-end using a differentiable probabilistic PnP solver. This moves beyond traditional geometric consensus by explicitly modeling correspondence reliability, leading to better pose estimation. The learned confidence is then used to reweight flow supervision and guide correspondence selection at test time, resulting in significant accuracy gains over baseline methods on M3ED and DSEC datasets.