Adding curvature-based energy regularization to neural implicit representations produces smoother, artifact-free 3D maps from noisy LiDAR data while maintaining geometric accuracy and keeping inference fast.
This paper improves neural implicit representations for LiDAR mapping by adding geometric smoothness constraints. Instead of just fitting noisy point clouds, the method regularizes the neural signed distance field using classical surface energy functionals based on curvature, resulting in cleaner 3D reconstructions without slowing down inference.