Dual Double Point Cloud Transformer

Long-range context relationship plays an important role in understanding point clouds. Inspired by the permutation invariance of the recent Transformer models, in this paper, we propose an end-to-end Dual Double Transformer (DDTFormer) architecture for 3D point cloud processing. Specifically, two core components, termed point-wise double transformer and channel-wise double transformer blocks, are well designed for explicitly modeling interdependencies among points and channels. And both blocks consist of two key operations: aggregation attention mapping the point-wise/channel-wise feature maps into a global space, and propagation attention diffusing the aggregated features back to the input points or channels. We illustrate the effectiveness and competitiveness via extensive quantitative and qualitative experiments on 3D shape classification and segmentation benchmark datasets.