A graph cut optimization guided by 3D-features for surface height recovery

This paper aims to present a new approach for automatic urban scene modeling from high-resolution satellite images with focus on building areas. The input data consist of a panchromatic stereo pair of satellite images, with a submetric resolution of 50–70 cm and a low Base to Height ratio B/HB/H [0.05–0.2]. Since a detailed extraction and description of building roofs is complex in a satellite context, we propose to describe the scene by means of a 3D-surface that provides either raster or vector information using different description levels. The main contribution of our approach is the use of 3D-features such as 3D-segments and 3D-facets to guide the optimization process. 3D surface modeling can be formulated as a matching problem that can be solved by graph cut minimization. The novelty consists in the original construction of the graph to combine input 2D data and 3D feature constraints to control the final surface. Complementary features are used. 3D-segments modelize discontinuities and 3D-facets help to regularize the surface by planar patches. The proposed automatic system provides a surface height map with subpixellar precision. Moreover, the system is generic and extensible to other data such as aerial and terrestrial images or to a multiple view context. External databases can also be easily added to the process to constrain the optimization.