3D Point Clouds Classification and Segmentation Network Based on Global Features

Abstract

Point cloud classification and segmentation are the primary tasks in 3D computer vision with great application value. Recently, several methods adopt deep neural networks to solve the problems by directly taking the point clouds as input due to their simplicity and effectiveness. However, existing methods only focus on extracting local information of the point cloud and ignores the global features which carry information of great importance in 3D geometric space. Inspired by the way humans observe a 3D object (analyzing its 3D overall characteristics first and then combining it with more specific details), we propose a global feature-based dynamic graph convolutional neural network (GF-DGCNN) to simultaneously fuse global and local features of point clouds. Firstly, the feature extraction module is used to calculate global features of the point clouds, and then the local geometric features of the point clouds are extracted by edge convolution. Finally, the accuracy of the classification and segmentation problem is improved by fusing the global features and local features of the point clouds. Experimental results show that the proposed GF-DGCNN achieves state-of-the-art performance on standard benchmarks, including ModelNet40 and ShapeNet.