Improvement of Weakly Supervised Building Region Extraction Using Graph-Based Segmentation

In this study, we propose a method that can produce pseudo-pixel-wise labels from bounding box annotations to improve weakly supervised building region extraction using deep neural networks (DNNs). The proposed method aims to initialise GrabCut, which can refine a building and its background region based on graph theory, accurately using knowledge of the shape and distribution of buildings in the image. An experiment showed that the proposed method significantly increased the accuracy, which was measured by the Dice score, of the pseudo-labels by 1.46% compared to a conventional method. This result is supported by the fact that the accuracy of the GrabCut initialisation was 2.81% higher. In addition, the proposed method was effective because the accuracy of the weakly supervised DNNs increased significantly, by 1.91% or more. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.