Unsupervised arbitrary-scale point cloud upsampling by learning neural gradient function Unsupervised arbitrary-scale point cloud upsampling by learning neural gradient function

• Published in: Multimedia systems Vol. 31; no. 4
• Main Authors: Gao, Tao, Feng, Jiangshan, Wu, Xiaoqun, Li, Haisheng, Wang, Xiaochuan
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.08.2025; Springer Nature B.V
• Subjects: Computer Communication Networks; Computer Graphics; Computer Science; Cryptology; Data Storage Representation; Image segmentation; Multimedia Information Systems; Operating Systems; Regular Paper; Semantic segmentation; Point cloud upsampling; Density-sensitive; Neural gradient function; Arbitrary scale; Unsupervised learning
• ISSN: 0942-4962; 1432-1882
• DOI: 10.1007/s00530-025-01870-x

Point cloud upsampling aims to generate a dense and uniform point cloud from a sparse input, supporting various downstream tasks such as surface reconstruction and semantic segmentation. Current point cloud upsampling approaches mainly rely on ground truth complete point clouds as supervision, which are rarely available in real-world situations. Moreover, they are subject to fixed upsampling scale, thus are inconvenient to produce dense point clouds with desired resolution. To address these issues, we propose an arbitrary-scale point cloud upsampling method without supervision. Firstly, we employ a density-sensitive interpolation strategy designed to restore the sparse input point cloud into a dense point cloud of arbitrary scale. Then we derive the neural gradient function from the input point cloud to optimize the locations of surface points, thereby reducing noise and outliers. Experimental results demonstrate that our method not only achieves better visual results with fewer parameters but also exhibits competitive performance on the PU-GAN and PU1K datasets.