||Rumpler Markus, Irschara Arnold , Wendel Andreas, Bischof Horst
||Proceedings of the 17th Computer Vision Winter Workshop (CVWW)
||We present a fully automatic, object-centered approach for creating 3D city models by combining public domain geographic information with highly overlapping aerial images. In a first step, we create a polygonal 3D surface mesh from terrain elevation data provided by the Shuttle Radar Topography Mission. Second, we transfer building outline polygons from 2D vector maps provided by the OpenStreetMap project to georeferenced aerial images and extract image patches of individual buildings. Subsequently, we perform dense multi-view image matching to estimate a depth map for the roof pixels from which we approximate building heights and roof structures. This allows to reconstruct a polyhedral 3D city model of an area of 1 square-km with more than 700 buildings in less than 10 minutes on commodity hardware. Our model satisfies the CityGML Level-of-Detail accuracy requirements of +/- 2 m, evaluated by comparison to reference height data from airborne laser scanning. The obtained textured meshes are suitable for city planning, interactive visualization or simulations and may serve as base meshes for future refinement of the geometry.