APUNet: Attention-guided upsampling network for sparse and non-uniform point cloud

Abstract

Point cloud upsampling is a basic low-level task, that is important for improving the quality of a point cloud. However, existing point cloud upsampling methods perform poorly on sparse and non-uniform point clouds, due to that they fail to fully model the relationship between points. To address this issue, in this paper, we propose an attention-guided network called APUNet to exploit the correlation between points, which can perform unsampling for sparse and non-uniform point cloud. In particular, we first propose a feature extraction unit, DisTransformer, which can effectively model the relationship between points by introducing a distance prior to the attention mechanism. We also design a point cloud feature extraction network based on DisTransformer. By computing the correlation between patches and the correlation between points, we fuse the global and local features to better model the correlation of the whole object. Furthermore, we propose a feature prediction module based on attention mechanisms that avoids generating clustered points by transforming the point cloud expansion task into a point cloud prediction task. Qualitative and quantitative experiments reveal the superiority of our method compared to the current state-of-the-art methods. Compared with other point cloud upsampling methods, APUNet can much better upsample non-uniform and extremely sparse point clouds.