PU-RRN: Point Cloud Upsampling with Reversible Residual Network

Point cloud Upsampling (PU) improves the performance in downstream tasks by enriching the representation of raw point clouds. Most learning-based point cloud upsampling methods follow a pipeline of feature extraction, feature upsampling, and 3D coordinate reconstruction. However, their performance is limited by the reconstruction process of mapping back from the feature space to the 3D space. To address this issue, we propose a novel point cloud upsampling network called PU-RRN. Our method establishes a bidirectional mapping between the feature space and the 3D space through Reversible Residual Network, which can significantly reduce the accuracy loss during the reconstruction process. For feature extraction, we combine global and local features to better preserve the structural information of the original point cloud. Quantitative and qualitative evaluations demonstrate that our method achieves improvements in performance.