Block-based depth estimation from image triples with unrestricted camera setup

A block-based depth estimation approach is presented that evaluates image triples captured with an unrestricted calibrated camera setup. From each image triple, three image pairs are formed which are evaluated independently after respective rectification. The two resulting depth maps per image, one with respect to each of the other images, are averaged according to their reliability. Compared to binocular depth estimation, the presented trinocular approach reduces the depth error variance by 40 % and increases the number of reliable depth estimates significantly. In the European ACTS PANORAMA project, it is sucessfully applied to 3-D reconstruction of moving objects for video communications with 3-D telepresence