Point cloud upsampling via implicit shape priors discovery and refinement

Abstract

The point clouds obtained by scanning sensors are often sparse and non-uniform, therefore, point cloud upsampling is of vital importance. This paper considers geometric priors as a rich source to guide point cloud generation for the better qualities. However, it is less flexible to explicitly exploit geometric priors of object surface, such as local geometric smoothness and fairness. In light of this, this paper proposes a novel two-stage method via discovering and exploiting implicit shape priors, which can consist of coarse point cloud upsampling and fine details refining. Specifically, at the first stage, we explore to discover geometric priors in an implicit manner via Dual Transformer, which simultaneously addressing local and global information during feature encoding, while a Neighborhood Refinement module is proposed to handle with geometric irregularities and noises via exploiting feature similarity of neighboring points. Extensive experiments on synthetic and real datasets validate our motivation, demonstrating that our method achieves competitive performance compared to SOTA methods, and better results for noisy point clouds. The source code of this work is available at https://github.com/Vencoders/PU-DT.