PointPAVGG: An Incremental Algorithm for Extraction of Points’ Positional Feature Using VGG on Point Clouds

Abstract

Many works have been devoted to improving the accuracy of point cloud classification and segmentation, which are essential problems in computer vision. Although these works have achieved excellent performance using advanced feature extraction methods, it is still a challenging task to extract high-level features from disordered data in the form of point clouds. To tackle this issue, we propose a VGG-based network, called Point Positional Attention VGG (PointPAVGG) for 3D point cloud feature extracting and processing, which is inspired by the classical VGG network. Concretely, in order to combine global and local features, we extract the local point cloud geometric information by every sphere domain and analyze its global position score by our point attention (PA) module. This novel network, namely, PointPAVGG, with graph structure point cloud feature extraction and PA, is mainly presented and applied in point cloud classification as well as segmentation tasks. Comprehensive experiments carried out on ShapeNet and modelNet, which demonstrate that our methods deliver superior performance, showing state-of-the-art results in classification and segmentation tasks.