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

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