We present a motion estimation algorithm that tries to fit the best motion model to a given pair of images. Q-warping corresponds each point in one image to one point in another image according to some quadratic surface and some camera motion. This can be seen in this figure. We present real-life results that demonstrate the applicability of the method.
Among all possible quadrics and all possible camera motion, there are ones that describe the real motion best. Our algorithm searches for them. The estimation is done automatically and uses direct methods.