Dual Attention Network for Point Cloud Classification and Segmentation

Point cloud analysis has attracted increasing attention in recent years due to its wide applications in autonomous driving and robotics. However, how to learn distinct and informative features for point clouds is remaining challenging due to the disorder, irregularity, and sparsity of point clouds. In this paper, we propose a new network named Dual Attention Network (DANet) for point cloud classification and segmentation. The proposed DANet mainly consists of two modules, a local feature extraction module (LFE) and a global feature fusion module (GFF). The LFE enhances the learned local features by using the explicit geometric structure and implicit feature information, while GFF aims to fuse global information by self-attention. To demonstrate the effectiveness of DANet, we conduct experiments on two point cloud datasets, ModelNet40 for classification and ShapeNet Part for segmentation. Compared with the state-of-the-art methods, our network shows superiority on these point cloud tasks.