Coupled autoencoders learning for zero-shot classification with domain shift

Abstract

Zero-shot classification (ZSC) aims to classify images from the class whose training samples are unavailable. A typical method addressing this issue is to learn a projection from feature space to attribute space so that a relation of training samples and test samples could be built. However, the projection merely learned from training samples does not apply in unseen classes due to domain shift between them. To tackle this issue, we propose a novel method in this paper that jointly learns coupled autoencoders to alleviate the distribution divergence of samples. We learn a projection by adopting encoder-decoder paradigm in both seen and unseen classes. The proposed method is evaluated for zero-shot recognition on two benchmark datasets, achieving competitive results.