A Hybrid Method for Window Detection on High Resolution Facade Images

In this paper we present a hybrid method for detecting windows on high-resolution rectified images of building facades combining deep learning with traditional geometric processing. As initial step we use a deep learning object detection method. As we observed that in most cases the detector outputs a larger object than the ground truth. We employ geometric processing based on image gradients to precisely delineate the window edges. For the evaluation of the algorithm we have created a high resolution dataset with more than 2000 annotated windows. The obtained results show that the detector’s bounding box differs from ground truth mostly by less than six pixels. The Intersection over Union IoU of the objects is 96.9%. Geometric processing improves IoU by 1.7% leading to an IoU score of 98.6%.