Abstract

Image matching (or registration) is, in general terms, aligning two identical or similar images or parts of images, and it is basic to a number of vision problems. These include optical navigation (also known as motion analysis), stereo image interpretation, object analysis, change detection and others. The difficult part of all these problems, is the matching problem, i.e. the computation of the correspondence of elements in the different views. The computation of correspondence, in both the stereo and the motion problem, is impossible without the use of some assumptions about the visual world and various researchers have proposed several theories for the computation of visual correspondence that worked well for the domains that fitted their assumptions. In this paper we study the problem of the computation of visual correspondence, under the assumption that some values of the correspondence function are already known and we develop the mathematical constraints that are needed for growing a solution for the correspondence function from points where its value is known. In other words, since the problem of determining visual correspondence has proven to be very difficult, we attempt to solve a simplified version of it, i.e. the case where we already know the values of the correspondence for some points. This assumption is not very restrictive, since in several practical situations prominent features could be used to easily find the correspondence of some points.

© 1987 Optical Society of America