The authors present a novel approach to the correspondence problem and apply the resulting algorithm to the problem of stereopsis. Their algorithm approaches correspondence by providing an information-theoretic distance metric for the amount of information that a particular global match introduces. The algorithm then searches the set of all possible matches for the particular match that preserves the greatest amount of redundancy between the two sets of measurements. A process of nil mapping handles tokens that have no corresponding match in the other image.

The paper assumes a reasonable knowledge of information theory, although the authors do provide references to texts on the subject. It is well written and the authors present the material clearly. As a general technique for solving the correspondence problem, their method suffers from the need to know a priori the entropy associated with particular primitive matches. Another weakness is the need to prune the search tree in order to avoid the explosion associated with the enumeration of all possible matching functions. The examples the authors provide are quite trivial (binary images of straight rods) and do not represent a good test of the algorithm’s performance. Additional tests with real images would enhance the paper’s impact and show the technique’s promise.