A framework for object class recognition with no visual examples

Traditional approaches in object class recognition utilize a large number of labeled visual examples in order to train classifiers to recognize the category of an object in a test image. However, the need for a large number of training data makes the scalability of this approach problematic. In this paper, we explore the recently proposed paradigm of attribute based category recognition for object category recognition without using any visual examples. This goal is achieved by introducing a textual based attribute representation of an image and using these attributes for object categorization. We propose the Sparse Representations (SRs) framework to achieve training-free and highly scalable attribute prediction. We investigate different approaches in mapping the predicted attributes to object classes using Nearest Neighbors and Support Vector Machines. Experimental results suggest that the use of the SRs framework in conjunction with an appropriate Nearest Neighbors scheme can improve prediction accuracy at a much lower computational cost.