Abstract
Abstract. Building semantic segmentation is key to many applications relying on 3D modeling of city buildings such as urban planning or business intelligence. Recent works have shown great improvements in this area thanks to artificial intelligence, but even state of the art neural networks encounter difficulties to generalize to buildings that are different from the training dataset. 3D modeling applications also requires the elevation information often retrieved from a pair of High Resolution satellite images. In this article, we show that using both images of a stereo pair as inputs to a neural network trained for building semantic segmentation achieves better results than using a single view. Especially, stereo training gives a greater ability to generalize. We show that using neural networks designed for disparity estimation performs well for building semantic segmentation from a pair of satellite views in epipolar geometry. We also discuss how radiometry and disparity both affect the definition of what a building is depending on the multi-view network architecture.