Prototype Matters: Domain-Wise Prototype Induction for Zero-Shot Recognition

Zero-Shot Learning (ZSL) has achieved remarkable progress, yet most existing methods still rely heavily on predefined semantic embeddings as supervisory signals. Such semantics are often incomplete, inaccurate, and insufficiently discriminative, resulting in ambiguous decision boundaries and limited generalization to unseen domains. In practice, predefined semantics frequently conflict with diverse instance-level visual patterns, making it extremely challenging for models to both separate classes and decode rich attribute variations under limited parameterization. Furthermore, substantial discrepancies between semantic and visual structures hinder the learning of discriminative class-level representations, especially under domain shift. Although recent efforts aim to align visual and semantic spaces, they largely overlook the importance of explicitly inducing well-separated, domain-adaptive class-level representations, which we argue is crucial for effective zero-shot recognition. Motivated by these observations, we propose a novel Domain-wise Prototype Induction Network (DPIN) that focuses on learning discriminative and adaptively separable class-level prototypes instead of directly embedding or matching predefined semantics. DPIN learns to induce prototypes guided by limited semantic priors while dynamically adjusting inter-class distances according to domain variations. Specifically, we introduce a Semantic Node Refinement with Hierarchical Guidance module to progressively rectify predefined semantics, and an Attention-based Visual Node Construction with Semantic Prior module to derive visually discriminative nodes under class- and group-level supervision. Furthermore, a Meta-Graph Learning with Interactive Knowledge Infusion mechanism collaboratively refines semantic and visual nodes at both distribution and class levels, effectively narrowing the visual-semantic gap. Based on the learned meta-graph, an Adaptive Class-level Prototype Induction strategy generates domain-wise well-separated prototypes, enabling accurate instance-level inference even under severe domain shift. Extensive experiments on three standard ZSL benchmarks demonstrate the effectiveness of DPIN, achieving 4.3% and 2.5% average accuracy improvements over state-of-the-art ZSL methods under the same settings, respectively, and validating the effectiveness of prototype-centric modeling in ZSL.