Towards automatic 3D reconstruction of urban scenes from low-altitude aerial images

Abstract

We propose a methodology for reconstructing large-scale architectural scenes from low-altitude aerial images, in an efficient, accurate and fully automatic way. Towards this goal, we have developed an area-based segmentation technique, called colored watershed, that is particularly suited to segmenting objects with homogeneous photometric properties, which are typical of such scenes. This technique is now being combined with a dense-stereo method biased towards depth discontinuities near the edges of the segmented objects. In a final step, parametric models of these segmented objects are instantiated and directly adjusted to the multiple images available to generate a mixed surface and elevation map for each scene.