Integrating Iconic and Structured Matching

Several investigations 11, 16, 19{21] have recently been un-dertaken into object recognition based on matching image intensity neighborhoods rather than geometric matching of features extracted from the images. These projects have used small subwindows or complete image regions and matching has been based on the similarity of extracted descriptors to previously stored descriptors. One characteristic common to these approaches is the representation of objects as a whole, rather than as a structured ensemble. This paper describes an extension to these approaches wherein a set of related features recognized at an earlier iteration also contribute to the complete object recognition. The paper describes an iconic, or image-based, matching approach that incorporates an element of geometric matching and shows that use of the subfeatures improves matching eeciency, position accuracy and completeness.