• New VPFSS model simplifies training and avoids complex distillation • Local token extraction helps capture fine image details for labels • Shared features let labels reuse cues, improving recognition of new classes • Shares visual features across labels to boost performance on rare and unseen categories In real-world image understanding, images often contain multiple semantic concepts, including rare or previously unseen categories, posing significant challenges for multi-label classification. Multi-Label Zero-Shot Learning (ML-ZSL) aims to recognize unseen labels without requiring annotated training data. While recent approaches based on Vision-Language Pre-training (VLP) models have shown promise, they often struggle to preserve fine-grained local visual features that are critical for multi-label recognition. In this work, we present Visual Projector with Feature Space Sharing (VPFSS), a lightweight framework that adapts frozen VLP models for open-vocabulary multi-label classification. The proposed method consists of two components: a visual token extraction module that derives localized visual tokens from spatial features, and a feature space sharing mechanism that assigns tokens to labels through similarity maximization. By decoupling token extraction from label matching and avoiding knowledge distillation, the framework maintains a simple training pipeline while enabling reusable visual representations across labels. Experiments on NUS-WIDE and MS-COCO under both Zero-Shot Learning(ZSL) and Generalized Zero-Shot Learning(GZSL) settings demonstrate that VPFSS achieves consistent improvements over representative baselines. The results indicate that explicit modeling of localized tokens combined with cross-label feature sharing provides a practical and effective solution for open-vocabulary multi-label recognition.
Zhao et al. (Sun,) studied this question.