DRGNet: Dual-Relation Graph Network for point cloud analysis

Abstract

Recently point cloud analysis has attracted more and more attention. However, it is a challenging task because point clouds are irregular, sparse, and unordered. To accomplish that task, this paper proposes Dual Relation Convolution (DRConv) which utilizes both geometric relations and feature-level relations to effectively aggregate discriminative features. The geometric relations take the explicit geometric structures to establish the spatial connections in the local regions while the implicit feature-level relations are taken to capture the neighboring points with the same semantic properties. Based on our proposed DRConv, we construct a Dual-Relation Graph Network (DRGNet) for point cloud analysis. To capture long-range contextual information, our DRGNet searches for neighboring points in both 3D geometric space and feature space to effectively aggregate local and distant information. Furthermore, we propose a Channel Attention Block (CAB), which puts more emphasis on important feature channels and effectively captures global information, and can further improve the performance of point cloud segmentation. Extensive experiments on object classification, shape part segmentation, normal estimation, and semantic segmentation tasks demonstrate that our proposed methods can achieve superior performance.