Clustering of matches and verification step

For each individual match, we maintain during the recognition step an estimation of the associated transformation by fusing the transformations between confirming frames. To group matches belonging to the same (rigid) substructure, we run a very simple clustering algorithm on the associated transformations. Indeed, up to measurement errors, frames should undergo a similar transformation within a single substructure.

Each cluster is refined by an iterative closest neighbor technique where we enforce symmetry of the matches and verify their validity with a test.

[ Back to substructure matching with frames ]