LAG-Net: dual-attention mechanism to capture local and global context for large-scale point cloud semantic segmentation

Semantic segmentation of large-scale three-dimensional (3D) point clouds is critical for artificial-intelligence-based scene recognition and understanding. In this paper, we first design a local context encoding module (LCE) to capture local features by integrating geometric information and semantic features. Then, we introduce a dual-attention mechanism (DAM) to adaptively learn locally weighted features, thereby reducing feature ambiguity and enhancing the representation of local contextual information. In addition, we propose a global context encoding (GCE) module to capture long-range dependencies by incorporating point coordinates and neighborhood volume ratios. Built upon these modules, LAG-Net effectively integrates local and global contextual features by exploring spatial dependencies, thereby improving semantic segmentation performance in large-scale real-world scenes.