GNDAN: Graph Navigated Dual Attention Network for Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) aims to recognize unseen classes by transferring knowledge from seen classes.
• Existing ZSL methods often struggle to effectively integrate global and local visual features and their inter-region relationships.
• This limits the discriminative power of learned feature representations.

Purpose of the Study:

• To propose a novel Graph Navigated Dual Attention Network (GNDAN) for enhanced zero-shot learning.
• To address the limitations of existing methods in exploiting appearance relationship priors between local image regions.
• To jointly learn cooperative global and local features for more discriminative representations.

Main Methods:

• GNDAN utilizes a region-guided attention network (RAN) for learning discriminative local embeddings via soft spatial attention.
• A region-guided graph attention network (RGAT) incorporates global context by learning attribute-based region features and their relationships using graph attention.
• A self-calibration mechanism matches learned joint visual embeddings with semantic embeddings for final predictions.

Main Results:

• GNDAN effectively exploits appearance relationship priors between local image regions.
• The network jointly learns discriminative local embeddings and incorporates global context for explicit global embeddings.
• Experiments on three benchmark datasets show GNDAN outperforms state-of-the-art ZSL methods.

Conclusions:

• The proposed GNDAN significantly advances zero-shot learning by effectively integrating local and global features through a graph-guided dual attention mechanism.
• The method demonstrates superior performance in recognizing unseen classes.
• The availability of code and models facilitates further research and application.