PAT: Point Cloud Analysis with Local Filter Embedding in Transformer

3D point clouds can offer rich geometric content of information for the scenes. How to learn geometric features locally and globally has become a growing concern recently. This paper proposes a novel approach for Point cloud Analysis based on Transformer network, called PAT, which utilizes interactive-attention to capture multi-angle features in irregular point clouds for classification and segmentation. Specifically, the double-coordinate encoder is designed as local filter operations by mining neighborhood information from point coordinates and direction simultaneously in the bi-coordinate system. The proposed PAT for point cloud analysis can interactively perform local and global attention across different feature spaces. Validation was conducted on two public datasets and results show that the proposed PAT achieves good performance.