Learning similarity features of point clouds in Euclidean space and feature space

Abstract

We propose to perform feature similarity learning (FSL) module on raw point clouds. Many works aggregate local information based on Euclidean space neighbors. However, several nonlocal point pairings share features that are very similar. Based on this, we divide the FSL module into two sub-modules: 1) initial local feature aggregation (ILFA) module based on Euclidean space and 2) nonlocal feature aggregation (NLFA) module based on feature space. Specifically, ILFA module utilizes the Fast Point Feature Histogram (FPFH) descriptor to aggregate local neighboring information. Then, the NLFA module searches the feature space calculated by the ILFA module firstly for comparable features. By encoding feature similarity, the NLFA module learns and aggregates nonlocal information. Our method can more effectively capture local and global geometric features by combining the two submodules. We apply the FSL module to the DGCNN frame network and achieve state of the arts performance on benchmarks such as ModelN et40, ShapeN et, and S3DIS.