TGNet: Aggregating Geometric Features for 3D Point Cloud Processing

Point cloud analysis is a challenging task due to geometric information hidden in disordered, unstructured points. In this paper, we propose a new framework named Tree Graph Network (TGNet) for sampling, grouping, and aggregating geometric features of point clouds. Specifically, we build a graph called Tree Graph with several curves extending in different directions through explicit rules, and then aggregate the graph using the cross-attention mechanism block. In this way, we incorporate more geometric information into local features. We demonstrate the excellent performance of our model on standard benchmarks for several basic point cloud processing tasks such as classification, segmentation, and normal estimation. We also provide ablation studies and visualizations that aid in understanding our network.