Wang, TaoWu, JingJi, ZeLai, Yu-KunVangorp, PeterHunter, David2023-09-122023-09-122023978-3-03868-231-8https://doi.org/10.2312/cgvc.20231188https://diglib.eg.org:443/handle/10.2312/cgvc20231188We introduce a novel approach to the completion of 3D scenes, which is a practically important task as captured point clouds of 3D scenes tend to be incomplete due to limited sensor range and occlusion. We address this problem by utilising sparse convolutions, commonly used for recognition tasks, to this content generation task, which can well capture the spatial relationships while ensuring high efficiency, as only samples near the surface need to be processed. Moreover, traditional sparse convolutions only consider grid occupancies, which cannot accurately locate surface points, with unavoidable quantisation errors. Observing that local surface patches have common patterns, we propose to sample a Radial Basis Function (RBF) field within each grid which is then compactly represented using a Point Encoder-Decoder (PED) network. This further provides a compact and effective representation for 3D completion, and the decoded latent feature includes important information of the local area of the point cloud for more accurate, sub-voxel level completion. Extensive experiments demonstrate that our method outperforms state-of-the-art methods by a large margin.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies -> Shape representations; Point-based modelsComputing methodologiesShape representationsPointbased models10.2312/cgvc.2023118829-379 pages