Point Cloud Semantic Segmentation Based on Boundary Awareness and Multiscale Spatial Descriptors

3D point cloud semantic segmentation is essential for 3D environment perception and scene understanding. A key challenge lies in enhancing object boundary segmentation accuracy and capturing global contextual features efficiently. However, most existing methods optimize boundaries only through local feature enhancement or explicit boundary optimization, and struggle to comprehensively capture global information. Therefore, we propose BAMS-Net, a novel point cloud semantic segmentation network, integrating boundary awareness and multi-scale spatial descriptors. Our model innovatively fuses local feature enhancement with explicit boundary optimization in depth, enabling improvement in boundary delineation. Specifically, candidate boundary points are first determined through a boundary prediction module based on local neighborhood semantic consistency, and an adaptive weight learning mechanism is adopted to suppress the feature interference from boundary points in their neighborhoods. Then the boundary segmentation performance is further improved through a boundary optimization aggregation loss that combines local semantic boundary loss and adaptive aggregation loss. To effectively capture global context, an adaptive calibration module is used to reweight the multi-level local features and strengthen the critical information, then the enhanced local features of each layer are converted into multi-scale global descriptor through cross-level fusion, thereby efficiently capturing global contextual information. Experimental results on the S3DIS and SensatUrban datasets demonstrate that BAMS-Net extracts richer point cloud features, delivers precise object boundary segmentation, and achieves superior overall performance in point cloud semantic segmentation.