Edge-based feature extraction module for 3D point cloud shape classification

Abstract

Edge points represent the basic topological shape of an object, and the edge features of point clouds are very prominent geometric information, which play a very important role in the accuracy of object recognition. Considering that it is challenging to apply deep learning to edge detection of point clouds, we improved the edge extraction algorithm based on Angle Criterion (AC) to obtain edge feature points. In addition, a plug-and-play edge-based feature extraction module is designed to encourage the learning of edge features. An RNN structure branch is included in the module to enhance the feature extraction ability of the module. Edge-based feature extraction module can be integrated into some classical neural networks to form a novel framework, called PointEF. Experimental results show that the improved AC edge extraction algorithm is robust to noise and edge sharpness. Moreover, extensive experiments confirm the proposed module’s effectiveness and robustness to improve the performance of various networks on shape classification.