Generative Multi-Label Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-label zero-shot learning (ZSL) aims to classify images with multiple unseen categories.
• Current methods struggle with reliable attention map computation for unseen classes in multi-label ZSL.
• Generative Adversarial Networks (GANs) excel in single-label ZSL but multi-label feature synthesis remains unexplored.

Purpose of the Study:

• To develop a novel generative approach for multi-label zero-shot learning.
• To address the challenge of synthesizing multi-label features from class attribute embeddings.
• To investigate effective fusion strategies for multi-class information in ZSL.

Main Methods:

• Introduced attribute-level, feature-level, and cross-level fusion approaches for multi-label feature synthesis.
• Utilized GANs to generate multi-label visual features from class attribute embeddings.
• Proposed a cross-level fusion strategy as the core of the generative model.

Main Results:

• The proposed cross-level fusion-based generative approach achieves state-of-the-art performance.
• Outperformed existing methods on three major zero-shot benchmarks: NUS-WIDE, Open Images, and MS COCO.
• Demonstrated generalization capabilities in the zero-shot detection task on MS COCO.

Conclusions:

• This work presents the first generative approach for multi-label feature synthesis in zero-shot learning.
• The cross-level fusion strategy is highly effective for combining multi-class information.
• The method shows significant promise for advancing multi-label zero-shot image classification and detection.