Point cloud downsampling based on the transformer features

This research study delves into the issue of downsampling 3D point clouds, which involves reducing the number of points in a point cloud while maintaining high performance for subsequent applications. Current downsampling methods often neglect the geometric relationships among points during sampling. Drawing inspiration from advancements in the vision field, this paper introduces a point-based transformer to process point clouds with inherent permutation invariance. We have developed a transformer point sampling (TPS) module that possesses characteristics such as permutation invariance, task specificity, and noise insensitivity, making it an ideal solution for point cloud sampling. Experimental results demonstrate that TPS is effective in downsampling point clouds while capturing more detailed information, resulting in significant improvements for segmentation tasks.