Alleviating Feature Confusion for Generative Zero-shot Learning.

摘要: Lately, generative adversarial networks (GANs) have been successfully applied to zero-shot learning (ZSL) and achieved state-of-the-art performance. By synthesizing virtual unseen visual features, GAN-based methods convert the challenging ZSL task into a supervised problem. However, train generator on seen categories further apply it instances. An inevitable issue of such paradigm is that synthesized features are prone references incapable reflect novelty diversity real In nutshell, confusing. One cannot tell from ones using features. As result, too subtle be classified in generalized (GZSL) which involves both at test stage. this paper, we first introduce feature confusion issue. Then, propose new generating network, named alleviating GAN (AFC-GAN), challenge Specifically, present boundary loss maximizes decision ones. Furthermore, novel metric score (FCS) proposed quantify confusion. Extensive experiments five widely used datasets verify our method able outperform previous state-of-the-arts under GZSL protocols.