Multiscale Dual-Channel Attention Network for Point Cloud Analysis.

Point clouds are the most popular representation of 3D vision tasks and have a wide range of applications in the field of smart robots today. The disordered and unstructured nature of 3D points makes it more difficult for researchers to extract information from point clouds. When extracting information from point clouds, most methods ignore the geometric structure of local regions, making information extraction insufficient and thus affecting the model effect; or over-construct complex feature extractors to obtain more adequate information in local regions, which leads to extremely complex network models and dilutes the importance of points in the global structure. To this end, our approach proposes a multiscale dual-channel attention convolutional neural network model that considers both the extraction of information in the local structure and ensures the effectiveness of global information aggregation. The model effectively balances the fusion of local and global features and considers the effective combination of local and global features of point clouds in a more comprehensive way. It shows good performance on the classical datasets ModelNet40 and ShapeNet Part.