Learning Attribute Relation in Attribute-Based Zero-Shot Classification

Mingxia Liu, Songcan Chen, Daoqiang Zhang

Abstract: Recently, zero-shot learning has attracted increasing attention in computer vision community. One way of realizing zero-shot learning is by resorting to knowledge about attributes and object categories. Most existing attribute-centric approaches focus on attribute-class relation artificially derived by linguistic knowledge base or mutual information. In this paper, we aim to learn the attribute-attribute relation automatically and explicitly. Specifically, we propose to incorporate the attribute relation learning into attribute classifier design in a unified framework. Furthermore, we develop a new scheme for attribute-based zero-shot object classification, such that the learned attribute relation can be reused to boost the traditional attribute classifiers. Extensive experimental results demonstrate that our proposed method can enhance the performance of attribute prediction and zero-shot learning.