Hierarchical contrastive representation for zero shot learning

Zero-shot learning aims to identify unseen (novel) objects, using only labeled samples from seen (base) classes. Existing methods usually learn visual-semantic interactions or generate absent visual features of unseen classes to compensate for the data imbalance problem. However, existing methods ignore the representation quality of visual-semantic pairs, resulting in unsatisfactory alignment and prediction bias. To tackle these issues, we propose a Hierarchical Contrastive Representation learning paradigm, termed HCR, which fully exploits model representation capability and discriminative information. Specifically, we first propose a contrastive embedding, which preserves not only high quality representations but also discriminative enough information from class-level and instance-level supervision. Then, we introduce a regressor by valuable prior knowledge for conducting more desirable visual-semantic alignment for unseen classes. A pluggable calibrator is also aggregated to further alleviate prediction bias in contrastive embedding. Extensive experiments show that the proposed HCR can significantly outperform the state-of-the-arts on popular benchmarks under ZSL and challenging GZSL settings.