ACR-ILPU: Internal Learning 3D Point Cloud Upsampling Network Using Attention Coordinate Refinement

The performance of conventional externally trained networks for point cloud upsampling is often influenced by the semantic range size of the training dataset, leading to limited capability in learning shapes at all levels. Consequently, the upsampling of complex shapes fails to preserve their original characteristics and generates an excessive number of dense point cloud outliers. To address these issues, we propose an internally learned method for point cloud upsampling, termed Internal learning 3D Point Cloud Upsampling Network using Attention Coordinate Refinement (ACR-ILPU). ACR-ILPU comprises three modules: an edge-aware generator, a discriminator, and attention-based coordinate refiners, which collectively facilitate the training and testing processes of the network. Notably, our approach eliminates the need for an external training dataset and solely relies on an input sparse point cloud for internal learning. To evaluate the effectiveness of ACR-ILPU, we conduct comprehensive qualitative and quantitative experiments using the PU1K dataset. The results demonstrate that ACR-ILPU significantly reduces training time while surpassing the performance of both externally trained and internally learned methods. Furthermore, ACR-ILPU exhibits promising results when applied to point clouds from the extensive KITTI autonomous driving dataset.