DRGCNN: Dynamic region graph convolutional neural network for point clouds

Abstract

Convolutional Neural Network (CNN) is good at processing regular data, but point clouds are regular and discretely distributed in space. In order to handle point cloud, we use graph to relate points to points and use graph convolution to process point cloud. In neurology, it is well known that the receptive field size of visual cortical neurons is regulated by stimulation. Different sizes of receptive fields have different information. So, we construct different dynamic graph with point’s neighbors to expand each point’s receptive field and adaptively select the information in the different size of receptive field. So in this paper, we propose a new convolution method, Dynamic Region Graph Convolution. It consists of three parts: graph construction, region selection, feature fusion. And it is easily integrated into other popular networks. In experiment, to proof the effectiveness and robustness, we choose different k to construct graph and different number of points for training. We show the encouraging performance on point cloud data sets ModelNet40, ShapeNetPart with 93.3% and 86.2%.