Simple and Effective Spatial-Attended Graph CNN with Learnable Aggregation for Classification in Point Clouds

The main task of point cloud classification is to extract unique features from each point cloud and to accurately distinguish individual point clouds. How to extract meaningful features in point clouds to improve recognition accuracy has always been a challenge. In this work, the double-KNN dynamic graphs with different neighbor sets are proposed to process the input point cloud separately and to extract respective features. These two separate features are individually gone through the convolutional block attention module (CBAM) [1], followed by learnable aggression mechanisms to construct the final salient features for classification. The simulation shows that the proposed architecture has better performance in overall accuracy and average accuracy when testing in the ModelNet40 dataset.