3D window localization on building facades from aerial images

Buildings, the key elements of urban 3D modeling and monitoring or thermographic urban 3D map generation, are mainly represented by their facades. Geometric and semantic properties can directly be derived when knowing a facade's windows locations and dimensions. We present a novel 3D window detection framework where due to the high intra-class variability and the low distinctiveness of windows, the focus of this paper is to not optimize the detector itself, but to exploit redundant detection candidates coming from multiple viewpoints. After detecting the candidates in each image individually and clustering them in 3D space, a Markov Random Field and geometric reasoning are exploited to analyze each cluster and to find out correct window candidates. Experiments show the practicability of the proposed approach which outperforms state-of-the-art 2D detectors on a variety of datasets.