Point Cloud Completion with Difference-Aware Point Voting

Lihua Lu, Ruyang Li, Hui Wei, Yaqian Zhao, Rengang Li, Binqiang Wang

Abstract: Previous methods usually recover the complete point cloud according to the shared geometric information between the known and unknown shapes, which neglects that the partially known points and the points to be inferred may be different referring to both spatial and semantic aspects. To address this problem, we propose a model with Difference-aware Point Voting to reason about the differences between the known and unknown point clouds. Before voting, we devise a Multi-scale Feature Extraction block to filter out low-quality observed points obtaining the seed points, and learn each seed point with multi-scale features facilitating the voting process. By voting, we fill in the gaps between the known and unknown shapes in feature space, and infer the features to represent the missing shapes. Further, we propose a hierarchical point decoder to progressively refine the voting process under the guidance of the geometric commonalities shared by the observed and missing parts. The decoder finally generates the new points that can be taken as the center points for generating missing regions. Quantitative and visual results on PCN and ShapeNet-55 datasets show that our model outperforms the state-of-the-art methods.