ILPG: Instance-Level Prototype Generation for Zero-Shot Learning

xiaowen yan, Jiangnan Li, Fuyu Qi, Kaixi Wang, Jinfu Fan

Keywords: Instance-Level Prototype Generation;Zero-Shot Learning;Dynamic Instance Semantic Vectors;

Abstract: Zero-shot learning (ZSL) aims to recognize unseen classes by transferring knowledge from seen ones through shared semantic attributes. However, existing methods typically align image features with static, class-level prototypes, which ignore intra-class diversity, lack adaptivity to individual samples, and often suffer from semantic drift. We propose the Instance-Level Prototype Generation (ILPG) network, a lightweight framework that dynamically refines semantic prototypes on a per-instance basis. ILPG combines an attention-based attribute localization module, which highlights discriminative visual regions, with a semantic adjustment pathway that personalizes class prototypes to capture instance-specific variations. This design achieves fine-grained alignment between image features and class semantics while mitigating prototype rigidity. To further enhance robustness, we introduce a synergistic loss formulation that balances discriminability and semantic consistency, ensuring dynamically adjusted prototypes remain semantically faithful. Extensive experiments on three widely used benchmarks (CUB, SUN, and AWA2) demonstrate that ILPG consistently outperforms competitive baselines. ILPG not only establishes new state-of-the-art performance in both conventional and generalized ZSL but also provides interpretable attribute–feature associations.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 18605