Triple Verification Network for Generalized Zero-Shot Learning

Haofeng Zhang, Yang Long, Yu Guan, Ling Shao

Abstract: Conventional zero-shot learning approaches often suffer from severe performance degradation in the generalized zero-shot learning (GZSL) scenario, i.e., to recognize test images that are from both seen and unseen classes. This paper studies the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Class-level Over-fitting</i> (CO) and empirically shows its effects to GZSL. We then address ZSL as a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">triple verification</i> problem and propose a unified optimization of regression and compatibility functions, i.e., two main streams of existing ZSL approaches. The complementary losses mutually regularizes the same model to mitigate the CO problem. Furthermore, we implement a deep extension paradigm to linear models and significantly outperform state-of-the-art methods in both GZSL and ZSL scenarios on the four standard benchmarks.