A closed form solution for viewing graph construction in uncalibrated vision

This paper presents a closed form solution for the problem of computing a set of projective cameras from the fundamental matrices of a given viewing graph. The approach is incremental, exploits trifocal constraints, and does not rely on either image or structure points. Represented by a vector of four parameters that uniquely ensure its consistency with the local trifocal geometry, each newly computed camera is automatically coherent with the projective frame chosen as global reference, thus not needing any a posteriori synchronization. Results of experiments made under controlled conditions show that the proposed approach is relatively resilient to noise, and faster by three orders of magnitude than classical camera resectioning solutions, while reaching a comparable accuracy. This makes our closed form approach a good candidate for camera initialization in scenarios involving large-scale viewing graphs.