This is published under the terms of CC BY-NC.

Downloaded: 97 times

Abstract

Human point cloud completion is a challenging yet indispensable task, devoted to filling missing parts in the collected incomplete point clouds. Existing methods overly rely on features extracted from surface points, neglecting the intrinsic joints information point clouds possess. To address this problem, we propose a new network with an encoder-decoder framework, named JointFormer. Firstly, we design a joint-enhanced encoder that provides more prior guidance on the overall structure of the partial input. Then, a generator is employed to generate sparse but complete point clouds. Finally, a decoder refines the rough point clouds into complete and dense human body point clouds in a coarse-to-fine manner. Moreover, combining transformer with the Convolutional Block Attention Module (CBAM), we design the Channel-Spatial Attention Transformer (CSAT) to better capture point cloud spatial relationships. Quantitative and qualitative evaluations demonstrate that JointFormer outperforms the state-of-the-art completion method on our two human body point cloud datasets.