Edited by author 11-12-2002 12:57 PM
This paper definitely has some very nice ideas. I especially like the automatic recovery of the number of motions.

I do wonder, however, about the applicability of this approach. This work seems to require that ALL correspondences are correct. This is tricky for most scenes but even more difficult for objects that are not roughly planar - the type of objects that are needed to specify their own fundamental matrix. In other works do the author try to detect objects for which a homography would be a better model?