KC-PointNet: Attentional Network For 3D Point Cloud Processing

Abstract

Point cloud analysis has always been a challenging problem mainly due to the orderless nature of point cloud. PointNet uses point-wise MLP to solve this problem and achieves impressive results. But one limitation of PointNet is that they treat all the points and all the feature channels equally, this is unreasonable since there are always some key points and key feature channels are crucial to recognizing the object which should be paid more attention to. In order to overcome these two shortcomings, we propose the Key-point and Channel Attention (KCA) module, a simple but effective plug-and-play attention module for point cloud processing. In this module, point wise attention learn to select the key points, while channel wise attention learn to select more distinctive feature channels, which will make our features more representative. Our KCA module is also lightweight which introduces minor computation efforts.We evaluated our method on the public benchmark ModelNet40 and verified its effectiveness in classification tasks.