Joint probability estimation of attribute chain for zero-shot learning

Zero-shot learning (ZSL) aims to classify the objects without any training samples. Direct Attribute Prediction (DAP) gives a solution with attribute space but it makes the assumption of attribute independence. To relax this assumption and consider the relation among attributes, Joint Attribute Chain Prediction (JACP) algorithm is proposed in this paper. It estimates the joint probability of attribute chain by the multiplication formula without the independence assumption. To deal with the difficulty of estimation, clustering algorithm of attributes is presented to generate several independent attribute sets. Each set trains a series of classifiers to calculate the joint probability individually and then the posteriori of classes can be obtained. Experiments on AwA and aPascal-aYahoo data sets demonstrate the effectiveness of our algorithm.