The goal of this collaborative proposal between the University of California at Santa Cruz and Duke University is the definition of features for recognition and classification of objects in imagery. Feature vectors are enriched and made invariant at the same time by an efficient, probabilistic coding of the feature-space manifolds that are obtained when features undergo all possible transformations in a predefined class. This coding is designed to circumvent the curse of dimensionality by reducing the representation requirements to the two properties that are essential for recognition and classification, that is, uniqueness and continuity. Uniqueness requires different manifolds to be coded differently, while continuity requires similar codings for similar manifolds.
Theoretical and empirical investigations will study optimal feature design, the performance of these features with various classification and recognition methods, and the trade-off of computational efficiency and performance. Misclassification rates for families of features and transformations will also be measured, both by theoretical bounds and empirical tests. The behavior of the proposed representations will be examined in the presence of clutter and data corruptions such as occlusions. These concepts will be tested through their application to the automatic diagnosis of colon cancer from computerized tomography scans, and to the interpretation of American sign language from video sequences.
