SDANet: spatial deep attention-based for point cloud classification and segmentation

Using deep learning to learn point cloud features directly have become one of the research hotspots in the field of 3D point cloud processing. The existing methods usually construct local regions, extract features from local regions, and then aggregate global features through multi-layer perceptron and maximum pooling layer. However, most of these processes do not consider the contribution of point cloud local features to the final decision and the spatial relationship between neighbor points, which limits the accuracy of 3D point cloud classification and segmentation. In this article, a novel network model called spatial depth attention network is designed to improve the accuracy of point cloud classification and segmentation, which embeds local depth attention mechanism into MLP layer to learn local neighborhood geometric representation. The local deep attention of the point cloud is obtained through the SDA module, and then combined with feature learning and local deep attention to effectively capture the local geometric structure. In order to achieve the best feature extraction ability, local depth attention features are combined with global features. Experiments show that SDANet achieves the same or better performance as the most advanced methods on several challenging benchmark datasets and tasks.