Geometry-Aware Transformer for Weakly Supervised Semantic Segmentation of Aerial Point Clouds

Aerial point cloud semantic segmentation technology enables intelligent interpretation of three-dimensional geographic information with precise structural representations and surface features. Recent Transformer-based point cloud semantic segmentation methods using multi-layer self-attention mechanisms can dynamically learn important relationships between points. With sufficient labeled samples, it can effectively capture complex global structures and long-range dependencies within point clouds. However, point-wise annotations of large-scale aerial point clouds are labor-intensive and time-consuming. Under limited annotation conditions, there are critical limitations of Transformer-based point cloud semantic segmentation methods, including inadequate modeling of local structures and ineffective classification of small objects. Therefore, to address the aforementioned challenges, this study proposes a geometric-aware transformer for a weakly supervised semantic segmentation method for aerial point clouds. Firstly, a geometry-aware encoding layer that explicitly incorporates foundational geometric descriptors (linearity, planarity, etc.) to compensate for the lack of fine-grained local supervision, enhancing feature discrimination under sparse labels. And a bidirectional interactive attention module that explicitly models both the influence of local neighborhoods on their center and the subsequent diffusion of global-context-aware semantics from the center back to its neighbors, thereby maximizing the utility of sparse annotations. Secondly, a pseudo-label adaptive training module leveraging abundant unlabeled data is designed to enhance model performance and generalization capability. Experimental results on the SensatUrban and DALES datasets demonstrate that the proposed method achieves 58.69% mIoU on SensatUrban (with 10% labeled data) and 74.80% mIoU on DALES (with 10% labeled data). These results surpass both the fully supervised Point Transformer baseline and multiple early fully supervised methods on both datasets.