'Walking on two legs': alternate use of two representations of a match during object recognition and location

Performance is presented for an object recognition and location method which alternately optimizes two representations of the match between object and image. The two representations are a geometric transform mapping the model onto the image coordinate and a feature pair list indicating the association of matched model/image features. Each optimization process has been designed to be history-free and consequently the system is capable of correcting mistakes formed in an earlier phase when the matching was rough.<
>