Instance-aware Contour Learning for Vectorized Building Extraction from Remote Sensing Imagery

Extraction of vectorized building instances from remote sensing images has made significant progress in recent years. The extraction of vectorized building maps enables rapid, large-scale updates to geospatial databases. Recently, the contourbased methods have achieved amazing success because their learning goals are more compatible with vectorized polygons. However, existing contour-based approaches achieve contour updates only by learning the local features at each vertex, which can lead to a contour smoothing problem. Besides, these methods achieve classification via features at the center point, which hinders the receptive field at the instance level. Our primary motivation is to overcome existing limitations introduced by local modeling during contour regression while simultaneously improving the instance classification performance. In this paper, we utilize instance-level features to guide the contour learning process. An Instance-Aware Contour Regression (IACR) module is designed to update the local contour features via crossattention on the instance-level features. Furthermore, based on the IACR module, we propose a novel vectorized building extraction framework BuildingVec, which interacts between the contour regression branch and instance classification branch through a well-designed cascade architecture. We also build a vectorized building dataset with fine-grained categories, named UBCv2 (vec), to benefit the study of vectorized extraction for fine-grained buildings. Experiments on the UBCv2 (vec) dataset demonstrate that BuildingVec achieves state-of-the-art performance compared to both mask-based methods and contour-based methods. When compared to other vectorized extraction methods on the Crowd AI dataset, BuildingVec also achieves a state-ofthe- art performance with an $\text{AP}_{50}$ of 93.1 %.