Shape Completion with Meso-Skeleton Learning

2021.4-2021.8

Point cloud completion is a significant topic in computer vision. Many deep-learning-based methods have been proposed to solve this problem directly based on an encoderdecoder structure. However, these architectures heavily rely on the representation ability of the encoded global feature. Some researchers try to leverage Meso-Skeleton to explicitly learn the global structure of objects. In this project we propose two architectures respectively backboned by PUNet and PF-Net to study the effect of Meso-Skeleton in improving point cloud completion. Experiments show that a high-quality skeleton largely boosts the shape completion performance in both CD and EMD scores.

[Project report] [code] [poster]