PST-NET: Point Cloud Sampling via Point-Based Transformer

Sampling is widely used for point cloud processing tasks, especially in autonomous driving domain with multiple 3D sensors to gather extensive point sets. However, geometric relations among points are rarely considered in sampling. Inspired by the recent advances in vision domain, the point-based transformer is introduced to process the point cloud with the inherent permutation invariance characteristic. We develop Point Sampling Transformer Network (PST-NET), including data augmentation, self-attention and local feature extraction, to generate optimal resampling distribution that is excellent for a particular point cloud application. PST-NET with characteristics of permutationin-variant, task-specific and noise-insensitive, is thus exactly suitable for point cloud sampling. Experiments verified that PST-NET successfully downsamples point cloud and captures more detailed information, with remarkable improvement for shape classification. Also various combinations of relation functions for self-attention are analyzed based on controlled experiments. The result shows that concatenation is more suitable for self-attention in sampling.