Fast optimal pose estimation for matching in two dimensions

The problem of searching for the occurrence of a model in a scene is a familiar one in object recognition, and has been extensively researched. The model may be present at any pose (i.e. may be translated, scaled and rotated), and may be obscured by clutter in the scene. We consider the search to take place after the model and the scene have been converted to sets of features. The search is then for a pose of the model which produces a close juxtaposition of a significant subset of its features to a subset of the scene features. In this process, trial sets of matches of pairs of model and scene features are assembled, and their geometrical consistency is tested. It is desirable, then, to find a procedure that will check two or more match pairs, produce a meaningful error estimate, will economically handle many match pairs and the incremental addition of new pairs to the set, will incorporate a realistic error function for a useful range of features, and will estimate an optimal pose for a given match set. The paper describes such a procedure for edge features in two dimensions