Self-supervised learning of pseudo classes for generalized zero-shot fine-grained recognition

Generalized zero-shot learning (GZSL) attempts to recognize visual instances from both seen and unseen classes by transferring knowledge from seen classes to unseen classes through semantic information (e.g., attributes). Generative methods are commonly employed to alleviate the issue of extreme data imbalance in which visual samples from unseen classes are not available during training, by synthesizing training samples for unseen classes from class prototypes. However, in the context of GZSL applied to fine-grained recognition, a notable complication arises. Similar class prototypes among different categories lead to ambiguity when generating synthetic data for classification. In response, we present a novel solution: a self-supervised pseudo-labeling (SSPL) module designed to enhance the generation of discerning synthetic data. This enhancement is achieved through an unsupervised grouping of fake and real samples using pseudo classes. By doing so, the SSPL module addresses the challenge of generating discriminative fake data, ultimately improving the overall quality of synthesized samples for classification. Our experimental results, conducted on three widely recognized GZSL datasets, demonstrate the effectiveness of the proposed method. Notably, the SSPL module not only produces well-distributed synthetic samples, but also enhances the discriminative and generalizable visual features derived from both real and synthetic samples within the GZSL framework.