Image Matching under Aane Deformations

Image deformations due to relative motion between an observer and an object may be used to infer 3-D structure. Up to rst order these deformations can be written in terms of an aane transform. Here, a novel approach is adopted to measuring aane transforms which correctly handles the problem of corresponding deformed patches. The patches are ltered using gaussians and derivatives of gaussians and the lters deformed according to the aane transform. The problem of nding the aane transform is therefore reduced to that of nding the appropriate deformed lter to use. The gaussian equations are solved both in the special case where the aane transform can be written as a similarity transform and for the general aane transform. The method is local and can handle arbitrarily large aane deformations. Experiments demonstrate that this technique can nd scale changes and optical ow in situations where other methods fail. The experiments also demonstrate the robustness of the technique.