GMP-Net: Graph based Missing Part Patching Network for Point Cloud Completion

Abstract

Estimating the missing part is an important branch of point cloud completion research. However, the existing point cloud completion methods ignore the importance of feature differences between partial and complete point clouds, and have not further optimized the predicted missing point cloud. For this reason, we propose an effective Graph based Missing Part Patching Network (GMP-Net), which is a deep learning model that uses the power of graph convolution to predict missing part of point cloud in a patching manner. First, GMP-Net improves its feature extraction ability through double perspective contrastive learning which extracts better global features with semantic boundaries between missing and incomplete point clouds. Second, a tree structure decoder combined with graph attention is proposed to generate intermediate coarse missing points by gathering topologies from ancestors’ information. Third, we propose a new graph convolution technique that refines intermediate generation by constructing a graph between partial and missing points. This approach optimizes the generated missing points with the local feature of the partial points. Extensive experiments demonstrate the effectiveness of the proposed GMP-Net.