An algorithm for optical ow estimation is presented for the case of discrete-time motion of an uncalibrated camera through a rigid world. Unlike traditional optical ow approaches that impose smoothness constraints on the ow eld, this algorithm assumes smoothness on the inverse depth map. The computation is based on di erential measurements and estimates are computed within a multi-scale decomposition. Thus, the method is able to operate properly with large displacements i.e., large velocities or low frame rates. Results are shown for a synthetic and a real sequence.
Jeffrey Mendelsohn, Eero P. Simoncelli, Ruzena Baj