Point clouds learning with attention-based graph convolution networks

Abstract

Point clouds, as a kind of 3D objects representation, are the most primitive outputs obtained by 3D sensors. Unlike 2D images, point clouds are disordered and unstructured. Hence the classification techniques such as the convolution neural network are not applicable to point cloud analysis directly. To solve this problem, we propose a novel network to extract point clouds feature, named attention-based graph convolutional network (AGCN). Taking the learning process as a message propagation between adjacent points, we specifically introduce attention mechanism to construct a point attention layer for analyzing the relationship between local points feature. The object classification is implemented by stacking multiple layers of point attention layer. In addition, the proposed network is extended to an attention-based encoder-decoder structure for segmentation tasks. We also introduce an additional global graph structure network to compensate for the relative location information of the individual points in the graph structure network. Experimental results show that our network has lower computational complexity and faster convergence speed. Compared with existing methods, the proposed network can achieve comparable performance in classification and segmentation tasks.