An Orthographic Mean-Shift Clustering and Segmentation Method for Building Mask Regularization

Convolutional neural networks often produce imperfect building masks that cannot be directly used to generate the regular polygons required in engineering applications. To address this issue, we propose a building mask regularization method that efficiently generates rectangular polygon building boundaries as post-processing. Our method represents the mask as a combination of rectangles by using a rectangular partition mechanism to separate the building mask into regular parts with boundaries close to bounding boxes. Each mask part is then rectified to align its longest edge with the horizontal line. Our proposed orthographic mean-shift clustering and segmentation groups rows and columns into refined building masks while removing unwanted noise such as seamless boundaries and burrs. Tests on two building datasets demonstrate that our method efficiently converts pixel-based segmentation masks into regularized building boundaries while preserving their shape. Furthermore, our method can produce better-regularized boundaries with higher intersection over union (IoU) scores for some low-quality building masks.