Deep Unbiased Embedding Transfer for Zero-shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot learning (ZSL) aims to classify objects not seen during training.
• Existing mapping-based ZSL methods struggle with projection domain shift due to limited training data.
• This domain shift hinders the accurate recognition of novel object categories.

Purpose of the Study:

• To propose a novel Deep Unbiased Embedding Transfer (DUET) model to mitigate projection domain shift in ZSL.
• To enhance the performance of zero-shot learning models on unseen classes.
• To introduce a quantitative metric for evaluating domain shift resistance.

Main Methods:

• Developed a Deep Unbiased Embedding Transfer (DUET) model comprising Deep Embedding Transfer (DET) and Unseen Visual Feature Generation (UVG) modules.
• The DET module integrates linear and nonlinear mapping functions for visual-semantic space connection.
• The UVG module employs a conditional generative adversarial network to synthesize visual features for unseen classes.

Main Results:

• The DUET model effectively alleviates projection domain shift in zero-shot learning.
• Experiments on five benchmarks demonstrate the superiority of the proposed DUET model.
• Both unseen class visual feature generation and the combined embedding transfer net contribute to improved ZSL performance.

Conclusions:

• The proposed DUET model offers a robust solution to the projection domain shift problem in zero-shot learning.
• End-to-end joint training and novel network architectures enhance model generalization.
• The developed ScoreRDS metric provides a valuable tool for evaluating domain shift resistance in ZSL models.