Dense Attention Network Based on Ordered Operation for Point Cloud Processing

Deep learning based 3D point cloud processing has become a new research hotspot. Existing methods either fail to fully exploit the three-dimensional characteristics of the points or ignore the connection among layers that are not directly connected. To solve these problems, this paper proposes a dense attention network based on Ordered operation for point cloud processing. We propose the O-edge module and a Residual Dual Channel Attention module (RDCAM) for point cloud feature extraction. The O-edge module is designed to extract edge features from ordered information by leveraging a series of one-dimensional convolutions, enabling better exploitation of the relations among points while retaining permutation invariant. RDCAM is used to draw network attention to key feature region. At the same time, we introduce the dense connection to point cloud processing to enhance information flow between layers. Experimental results show that the proposed network has higher classification accuracy than the existing networks on public benchmark of point cloud classification.