Causal Visual-semantic Correlation for Zero-shot Learning

Zero-Shot learning (ZSL) correlates visual samples and shared semantic information to transfer knowledge from seen classes to unseen classes. Existing methods typically establish visual-semantic correlation by aligning visual and semantic features, which are extracted from visual samples and semantic information, respectively. However, instance-level images, owing to singular observation perspectives and diverse individuals, cannot exactly match the comprehensive semantic information defined at the class level. Direct feature alignment imposes correlation between mismatched vision and semantics, resulting in spurious visual-semantic correlation. To address this, we propose a novel method termed Causal Visual-semantic Correlation (CVsC) to learn substantive visual-semantic correlation for ZSL. Specifically, we utilize a Visual Semantic Attention module to facilitate interaction between vision and semantics, thereby identifying attribute-related visual features. Furthermore, we design a Conditional Correlation Loss to properly utilize semantic information as supervision for establishing visual-semantic correlation. Moreover, we introduce counterfactual intervention applied to attribute-related visual features, and maximize their impact on semantic and target predictions to enhance substantive visual-semantic correlation. Extensive experiments conducted on three benchmark datasets (i.e., CUB, SUN, and AWA2) demonstrate that our CVSC outperforms existing state-of-the-art methods.