Attribute Based Classifiers for Image Understanding

Abstract

Attributes are mid-level semantic concepts which describe visual appearance, functional affordance or other human-understandable aspects of objects and scenes. In the recent years, several works have investigated the use of attributes to solve various computer vision problems. Examples include attribute based image retrieval, zero-shot learning of unseen object categories, part localization and face recognition.

This thesis proposes two novel attribute based approaches towards solving (i) top-down visual saliency estimation problem, and, (ii) unsupervised zero-shot object classification problem. For top-down saliency estimation, we propose a simple yet efficient approach based on Conditional Random Fields (CRFs), in which we use attribute classifier outputs as visual features. For zero-shot learning, we also propose a novel approach to solve unsupervised zero-shot object classification problem via attribute-class relationships. However, unlike other attribute-based approaches, we require attribute definitions only at training time, and require only the names of novel classes of interest at test time. Our detailed experimental results show that our methods perform on par with or better than the state-of-the-art.