A local–global approach to point cloud semantic segmentation

Point cloud semantic segmentation is pivotal for 3D scene understanding, yet it poses significant challenges in large-scale environments due to complex spatial structures and high inter-class similarity. This paper introduces the Local–Global Network (LG-Net), a symmetric encoder–decoder architecture designed to integrate local and global contextual cues through a dual-stage task-driven mechanism. The LG-Net features a Dual Attention Aggregation module for enhancing local boundary reasoning and a Global Semantic Relation module for capturing scene-level structure. Extensive experiments on the S3DIS and SensatUrban dataset demonstrate consistent accuracy improvements, validating the effectiveness of our approach for both indoor and outdoor 3D scenes.