Automatic window detection in facade images

摘要

City models play a major role in urban planning and are indispensable in nowadays civil engineering. The ongoing automation of simulations and analyses demand for increasingly detailed models. Especially windows are of high interest for several tasks. As city models commonly lack any relevant details, these have to be complemented by information about windows from other data sources. In this paper, we propose a pipeline to detect windows in ground view facade images which are rectified before detection. A postprocessing is applied to refine the detections made by a soft cascaded classifier and infer further windows. In experiments we compare our approach to previous work and evaluate the processing steps of our pipeline. Moreover, we show that our entire system yields a detection rate of 95% and a precision of 97% which is satisfying for a proper advancement of existing 3D city models.