Building Extraction From Satellite Images Using Mask R-CNN With Building Boundary Regularization

Kang Zhao, Jungwon Kang, Jaewook Jung, Gunho Sohn

Abstract: The DeepGlobe Building Extraction Challenge poses the problem of localizing all building polygons in the given satellite images. We can create polygons using an existing instance segmentation algorithm based on Mask R-CNN. However, polygons produced from instance segmentation have irregular shapes, which are far different from real building footprint boundaries and therefore cannot be directly applied to many cartographic and engineering applications. Hence, we present a method combining Mask R-CNN with building boundary regularization. Through the experiments, we find that the proposed method and Mask R-CNN achieve almost equivalent performance in terms of accuracy and completeness. However, compared to Mask R-CNN, our method produces better regularized polygons which are beneficial in many applications.