Rethinking Generalized Zero-Shot Learning: A Synthesized Per-Instance Attribute Perspective

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized zero-shot learning (GZSL) aims to improve model generalization to unseen classes.
• Existing GZSL methods struggle with the semantic gap and domain shift due to reliance on per-class attributes.
• Instance-level attributes offer a solution but require costly manual annotation.

Purpose of the Study:

• To propose a novel method, per-instance attribute synthesis (PIAS), for generating diverse semantic representations.
• To address the limitations of traditional GZSL approaches by eliminating the need for manual attribute annotation.
• To enhance the discriminability of visual and semantic representations for improved GZSL performance.

Main Methods:

• Utilizes Vision Transformer (ViT) for visual feature extraction and per-instance attribute generation.
• Defines class anchor points using generated attributes of class-average images and calibrates them in semantic space.
• Improves attribute diversity by aligning topological structures between annotated and synthesized attributes and features.

Main Results:

• PIAS significantly outperforms state-of-the-art methods on AWA2, CUB, and SUN datasets in both ZSL and GZSL settings.
• Demonstrates improved generalization capabilities of the proposed method.
• Successfully applied PIAS to attribute-based zero-shot image retrieval tasks.

Conclusions:

• PIAS offers an effective and efficient solution for generating diverse per-instance attributes in GZSL.
• The method successfully bridges the semantic gap and mitigates domain shift issues.
• PIAS shows strong potential for real-world applications requiring robust generalization to unseen classes.