Structure preserving point cloud completion and classification with coarse-to-fine information

Abstract

Point clouds are the predominant data structure for representing 3D shapes. However, captured point clouds are often partial due to practical constraints, necessitating point cloud completion. In this paper, we propose a novel deep network architecture that preserves the structure of available points while incorporating coarse-to-fine information to generate dense and consistent point clouds. Our network comprises three sub-networks: Coarse-to-Fine, Structure, and Tail. The Coarse-to-Fine sub-net extracts multi-scale features, while the Structure sub-net utilizes a stacked auto-encoder with weighted skip connections to preserve structural information. The fused features are then processed by the Tail sub-net to produce a dense point cloud. Additionally, we demonstrate the effectiveness of our structure-preserving approach in point cloud classification by proposing a classification architecture based on the Structure sub-net. Experimental results show that our method outperforms existing approaches in both tasks, highlighting the importance of preserving structural information and incorporating coarse-to-fine details.