Inferring Analogous Attributes: Large-Scale Transfer of Category-Specific Attribute Classifiers

The appearance of an attribute can vary considerably from class to class, causing standard class-independent attribute models to break down. Yet, training object-specific models for each attribute is impractical, and defeats the purpose of using attributes to bridge category boundaries. We propose a novel form of transfer learning that addresses this dilemma. Given a sparse set of class-specific attribute classifiers, our tensor factorization approach can infer new ones for object-attribute pairs unobserved during training. We apply our idea to learn over 25,000 analogous attribute classifiers on SUN and ImageNet.1