GSMFlow: Generation Shifts Mitigating Flow for Generalized Zero-Shot Learning

Generalized Zero-Shot Learning (GZSL) aims to recognize images not only for seen classes but also unseen ones by transferring semantic-visual relationships from the classes. It is an intuitive solution take advantage of generative models hallucinate realistic samples based on knowledge learned However, due generation shifts, synthesized most existing methods may drift real distribution data. To address this issue, we propose a novel flow-based framework that consists multiple conditional affine coupling layers learning data generation. Specifically, investigate and three essential problems trigger i.e., semantic inconsistency, variance collapse, structure disorder. First, improve reflection semantic information in generated samples, proactively embed into transformation each layer. Second, promote intrinsic feature variance classes, introduce boundary sample mining strategy with entropy maximization discover ambiguous visual variants prototypes hereby calibrate decision classifiers. Third, relative positioning proposed revise attribute embeddings, guiding which fully preserve inter-class geometric structure further avoid disorder space. Extensive experimental results four GZSL benchmark datasets demonstrate GSMFlow achieves state-of-the-art performance GZSL.