Abstract: Learning discriminative point-wise representations remains the central challenge in scene-level, correspondence-based point cloud registration. Most existing methods process two frames ...