Rectilinear Building Footprint Regularization Using Deep Learning

Abstract

Abstract. Nowadays, deep learning allows to automatically learn features from data. Buildings are one of the most important objects in urban environments. They are used in applications such as inputs to building reconstruction, disaster monitoring, city planing and environment modelling for autonomous driving. However, it is not enough to represent them in raster format, since applications require buildings as polygons. We use an existing, learning based approach to extract building footprints from ortho imagery and digital surface model (DSM) and propose a pipeline for building polygon extraction, which we call primary orientation learning (POL). The first step is to extract initial polygons, that contain a vertex for each pixel in the boundary of the footprint. Afterwards, the two primary orientation angles are regressed continuously. Using these orientation, we insert vertices such that all consecutive edges are perpendicular. To the best of our knowledge, our approach is the first to predict a continuous orientation angle for building boundary regularization. Furthermore, the proposed method is highly efficient with an average processing time of 2.879 ms for a single building.