A Graphical Convolutional Network-based Method for 3D Point Cloud Classification

Abstract

Point cloud data classification has been widely used in autonomous driving, robot perception, and virtual/augmented reality. Due to its irregularity and disorder, the classification task of point clouds needs to transform the point cloud into a multi-view or voxel grid, and then use the traditional convolution neural network processing. However, this process is not only complex in operation but also low in classification accuracy. To solve this problem, a new point cloud classification method based on the graphical convolutional neural network (GCN) is proposed. Firstly, based on PointNet, KNN graph is introduced to obtain global deep features. Then the 3D point cloud is represented as a directed graph, local features are extracted by edge convolution. Finally, the extracted global and local features are aggregated to represent and classify point clouds. The proposed network is evaluated on the open dataset ModelNet40 and 3DMNIST. Experimental results show that the proposed network can achieve on par or better performance than state-of-the-art, such as PointNet, PointNet++, DGCNN, and PointCNN, for point cloud classification.