PEAP-Net: Edge-Wise Attention Pooling for Point Cloud Classification and Segmentation

The irregularity and disorderliness of the point cloud data pose significant challenges for feature learning. Currently, most point cloud processing methods use symmetric pooling functions to aggregate local features. However, this approach may result in local point clouds being less sensitive to global features. In this paper, we develop an aggregated pooling module to construct a new network called PEAP-Net for point cloud classification and segmentation. The network primarily consists of two key modules: the Edge-wise Attention Pooling (EAP) block based on attention mechanisms and the Local Feature Aggregating (LFA) block based on graph convolutions. The EAP module incorporates the Edge-wise Attention structure, enabling efficient encoding of structural features for each input. To facilitate the integration of global features, we adopt a spatial attention strategy to fuse the input information. Additionally, we introduce a residual module to alleviate issues of information similarity and redundancy. The LFA module combines graph convolutions to group points and aggregate local features in the point cloud. We thoroughly evaluate the performance of the PEAP-Net model on two benchmark datasets: the ModelNet40 classification dataset and the ShapeNet segmentation dataset. Through numerous experiments, we demonstrate that our approach achieves significant improvements in both of these tasks.