This paper presents an algorithm for error detection and correction of disparity data as a process separate from stereo matching. The authors make the important point that the operation of matching is not necessarily the best place to exploit all the physical constraints characteristic of stereopsis. For example, if continuity is used as a criterion in matching, then it is lost as a tool for determining whether matches are correct.
Their method for error detection is somewhat limited in application because it requires matched points that are linked into edges. They further assume that the edges are straight lines with well-localized endpoints. This may be true of the Nevatia-Babu line finder but will not necessarily permit the extension of this method to matching results based on other edge finders.
In Section 4, “Qualitative Comparison of Matching Algorithms,” it appears that the authors do not use ground truth as a comparison. Rather, they assume that a rejection by their algorithm indicates that their match is in error, that is, they ignore the possibility that their algorithm can make mistakes too.
The paper is well written, and the authors acknowledge some of the failings of their technique. Specifically, their method is suitable for local errors but fails when more edges are matched to a wrong segment than to a correct segment. In addition, it cannot be used with area-based matching methods.