Justus H. Piater and Roderic A. Grupen, University of Massachusetts
A visual system that interacts with the real world must be able to adapt to unexpected situations and to flexibly discover relevant visual cues. We present a method that allows incremental learning of discriminative features. The feature space includes juxtapositions of k oriented local pieces of edge (edgels) and is parameterized by k and the relative angles and distances between the edgels. Specific features are learned by sampling candidate features from this space, increasing k as needed, and retaining discriminative features. For recognition of an unknown scene or object, features are queried one by one. As a result of each query, zero or more candidate object classes are ruled out that do not exhibit this feature to a sufficient degree. We discuss issues of computational complexity, and present experimental results on two databases of geometric objects.