GA-NET: Global Attention Network for Point Cloud Semantic Segmentation

Abstract

How to learn long-range dependencies from 3D point clouds is a challenging\nproblem in 3D point cloud analysis. Addressing this problem, we propose a\nglobal attention network for point cloud semantic segmentation, named as\nGA-Net, consisting of a point-independent global attention module and a\npoint-dependent global attention module for obtaining contextual information of\n3D point clouds in this paper. The point-independent global attention module\nsimply shares a global attention map for all 3D points. In the point-dependent\nglobal attention module, for each point, a novel random cross attention block\nusing only two randomly sampled subsets is exploited to learn the contextual\ninformation of all the points. Additionally, we design a novel point-adaptive\naggregation block to replace linear skip connection for aggregating more\ndiscriminate features. Extensive experimental results on three 3D public\ndatasets demonstrate that our method outperforms state-of-the-art methods in\nmost cases.\n