Regional features with adaptable global mappings for recognition systems

This paper proposes an adaptive recognition system that integrates local and global features while jointly classifying and learning from unlabeled data. Dictionaries based on local descriptors serve as the basis for the recognition system and at the same time provide spatial-mappings derived from the location of the selected features during classification, via l₁ minimization techniques. The mappings provide a global object representation that is utilized to discriminate among classes with candidate descriptors. Additionally updating or learning new local descriptors (via non-parametric Bayes) from unlabeled data within a dictionary framework, provides the flexibility needed when training data is limited.