Top-K attentional network for point cloud processing

Point cloud data is playing increasingly important roles in many applications such as autonomous driving, intelligent robots, virtual reality etc. Existing methods either have the problem of complex computation or fail to fully exploit the three-dimensional characteristics of the points. To solve these problems, this paper proposes a top-K attention network for point cloud processing. We propose the A-edge module for point cloud feature extraction which leverages the advantages of edge convolution and an attentional mechanism which enables the network focus on useful information. At the same time, top-K operator is introduced to replace the maximum pooling operation to retain more key feature information. Experimental results show that the proposed network has higher classification accuracy than the existing networks on public benchmark of point cloud classification task.