Employing hierarchical clustering and reinforcement learning for attribute-based zero-shot classification

Abstract

Zero-shot classification (ZSC) is a hot topic of computer vision. Because the training labels are totally different from the testing labels, ZSC cannot be dealt with by classical classifiers. Attribute-based classifier is a dominant solution for ZSC. It employs attribute annotations to bridge training labels and testing labels, making it able to realize ZSC. Classical attribute-based classifiers treat different attributes equally. However, the attributes contribute to classification unequally. In this paper, a novel attribute-based classifier for ZSC named HCRL is proposed. HCRL utilizes hierarchical clustering to obtain a hierarchy from the attribute annotations. Then the attribute annotations are decomposed into hierarchical rules which contain only a few attributes. The discriminative abilities of the rules reflect the significances of attributes to classification, but there are no training samples for evaluating the rules. The discriminative abilities are determined by reinforcement learning during the testing and the most discriminative rules are picked out for classification. Experiments conducted on 2 popular datasets for ZSC show the competitiveness of HCRL.