BuildMapper: A fully learnable framework for vectorized building contour extraction

Deep learning based methods have significantly boosted the study of automatic building extraction from remote sensing images. However, delineating vectorized and regular contours like a human does remains very challenging, due to difficulty methodology, diversity structures, imperfect imaging conditions. In this paper, we propose first end-to-end learnable contour framework, named BuildMapper, which can directly efficiently delineate polygons just as does. BuildMapper consists two main components: 1) initialization module that generates initial contours; 2) evolution performs both vertex deformation reduction, removes need for complex empirical post-processing used in existing methods. components, provide new ideas, including method replace methods, dynamic predicted ground truth pairing static correspondence problem, lightweight encoder information aggregation, benefit general contour-based method; well-designed classification head corner vertices detection, casts light on direct structured extraction. We also built suitable large-scale dataset, WHU-Mix (vector) The extensive experiments conducted WHU CrowdAI dataset verified achieve state-of-the-art performance, with higher mask average precision (AP) boundary AP than segmentation-based confirmed more 60.0/50.8% by test sets I/II, 84.2% set, 68.3% set are par manual delineation level.