TR-A-0149 :1992.8.19

Peter Meer

ERROR TOLERANT METHOD FOR INVARIANCE BASED FEATURE CORRESPONDENCE

Abstract:A two stage probabilistic algorithm for establishing feature correspondence between images is presented. First, k-tuples of features are matched based on similar invariant representations. The representation should be invariant not only to the group of image transformations but also to the permutation group of the k elements. A projective and permutation invariant representation for 5-tuples of points/lines in a plane is described. In the second stage, the algorithm recovers feature. correspondence from a contingency table built with the ensemble of matched k-tuples. The conditions for reliable performance are given, and it is shown that the correct solution for the correspondence problem can be be found while most the data in the matched k-tuples is erroneous. The algorithm does not use global information and recovers feature correspondence in O[n2] time.