Answers are grounded in this paper and cite their sources.
Building extraction is a critical task in the field of remote sensing. Many studies generate building polygons through segmentation followed by regularization, which often leads to cumulative errors and results in irregular shapes. The end-to-end methods generate building polygons directly from remote sensing images, improving overall accuracy. However, most existing methods rely solely on local features to generate building polygons, making it challenging to capture concave structures and often resulting in shape distortion. To solve this problem, this article presents an end-to-end building polygon extraction method referred to as global-aware building polygon extraction (GABPoly). Specifically, we design a ContourTransformer module with a stronger ability to capture global information. By learning the long-range dependencies between contour points, it ensures stable and consistent convergence during iterative updates. Furthermore, we introduce a graph-attention-driven vertex optimization method that leverages five geometric features and three edge relationships to mitigate the sample imbalance problem in vertex classification, thereby making the generated building polygons more regular and accurate. In addition, multiscale feature fusion is used to improve contour features, enabling the model to focus more effectively on building boundaries. Experiments on the WHU, Inria, and Vaihingen datasets demonstrate that GABPoly improves AP by 4.2, 3.5, and 16.5, respectively. Moreover, the method performs well on building polygons of different styles.
No takes yet. Share an insight, caveat, or question.
He et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: