Leclerc, Y. G. and Luong, Q.-T. and Fua, P. A Framework for Detecting Changes in Terrain, in Proceedings of the DARPA Image Understanding Workshop, Monterey, California, Nov 1998.
We propose a methodology that estimates the accuracy and reliability of the results of any multiple-image point correspondence algorithm, without the need for ground truth or camera calibration. The key concept behind our methodology is what we call the self-consistency of an algorithm across independent experimental trials (independent applications of the algorithm to subsets of images in a given collection of images). We describe precisely the self-consistency distributions and how to compute them. We then explain why we conjecture that they provide a reasonable approximation of the absolute error distributions. Experiments illustrate the usefulness of these distributions at ranking the quality of algorithms and the predictive power of scoring schemes.