K.U.Leuven > ESAT > PSI > Visics > Research > Topics > Item 1.9

3D city modelling

L. Van Gool, T. Moons, F. Mindru, K. Beets

Images of a city are taken as a car drives through the streets. From these images mosaics are produced of the facades of the different buildings. The production of these mosaics is highly automated. Invariant neighbourhoods are used to find corresponding features in subsequent images. These are small image patches that automatically adapt their shapes in the images in order to systematically cover corresponding, physical parts of a scene. The regions typically have a parallelogram or ellipse shape. The colour patterns within these neighbourhoods are characterized on the basis of moment invariants. Their values do not change under a series of geometric and photometric changes. This allows one to efficiently detect correspondences even if there are serious differences in viewpoint or illumination between the Subsequent images. Once correspondences have been found, appropriate projective transformations are applied to map the corresponding features onto each other and a photometric blending operation is carried out to build seamless mosaics. The partner in this IWT project is Eurosense NV.