CO-NET: Classification-Oriented Point Cloud Sampling via Informative Feature Learning and Non-Overlapped Local Adjustment

Abstract

Recent studies have proven the strength of task-oriented point cloud sampling methods over traditional non-learned ones. However, previous task-oriented samplers are not adequate to extract local details and spatial patterns of point clouds, limiting the quality of synthesized points. In this paper, we present a classification-oriented sampling network named CO-Net, aiming to learn informative sampled points that benefit downstream classification tasks. Specifically, we propose Attentive Point Spatial Pyramid (APSP) to achieve attention-based aggregation of multi-scale contexts. Then, we develop Gradient-weighted Feature Enhancement (GFE) to highlight class-discriminative structural details. We further design Non-overlapped Local Adjustment (NLA) to ease noise impacts with neighborhood contexts while preserving distinctive features of central points for more informative results. Experiments demonstrate that CO-Net outperforms state-of-the-art sampling methods.