Mohammad Rastegari

UWashington, student of Ali Farhadi

 

Abstract

Representing visual data with binary codes has been receiving great amount of attention recently. Binary codes not only enables fast search and retrieval, but also they boost the performance in recognition tasks. In this talk I will present the new notion of predictable binary codes that leads to the state-of-the-art object category retrieval. These binary codes can be seen as visual attributes in image domain. We extended the notion of predictability over textual domain and designed a dual-view model for learning the binary codes. These visual attributes are used to expand visual coverage of training set by adding images from a large pool of unlabeled images when one suffers from lack of training data in some categories.