Joint Feature Generation and Open-set Prototype Learning for generalized zero-shot open-set classification

Abstract

In generalized zero-shot classification, test samples can belong to either seen or unseen classes. However, in real-world situations, there may be many open-set samples in the test set where neither visual nor semantic representations of the classes are provided. The new problem is defined as generalized zero-shot open-set classification (GZSOSC). The purpose is to tell whether an instance belongs to which seen or unseen classes, or to reject an instance if it belongs to the open-set classes. To address this problem, we propose a novel method called Joint Feature Generation and Open-Set Prototype Learning (JFGOPL) for GZSOSC tasks. JFGOPL is presented to combine GAN training with open-set prototype learning, where the former generates high-quality unseen and open-set samples and the latter learns some open-set prototypes. Specifically, a novel GAN training strategy is proposed, where an intra-class compactness loss and an inter-class dispersion loss are proposed to ensure the discrimination of the generated samples and to make the learned embedding network less susceptible to the domain shift problem. Furthermore, open-set prototypes are derived by projecting confident open-set samples into the semantic space using the updated embedding network. Experiments on widely used benchmarks demonstrate the superiority of JFGOPL over existing methods for tackling the challenging GZSOSC problem.