CAT: Class-aware adaptive-thresholding for robust semi-supervised domain generalization

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain Generalization (DG) aims to transfer knowledge across domains, but requires extensive labeled data.
• High-quality labeled data is costly and labor-intensive, limiting practical DG applications.
• Semi-supervised Domain Generalization (SSDG) offers a label-efficient alternative.

Purpose of the Study:

• Investigate a practical SSDG problem under a label-efficient paradigm.
• Propose a novel method, CAT, for competitive generalization performance with limited labeled data.
• Address limitations of previous methods, including fixed thresholds and noisy pseudo-labels.

Main Methods:

• Leverage semi-supervised learning with limited labeled data.
• Employ adaptive thresholding for high-quality pseudo-label generation with class diversity.
• Utilize noisy label refinement techniques to enhance pseudo-label reliability.

Main Results:

• CAT achieves competitive generalization performance under domain shifts.
• Demonstrated superior performance on benchmark datasets: PACS (+3.45%), OfficeHome (+9.47%), and miniDomainNet (+10.90%).
• Highlights effectiveness in achieving robust generalization despite domain shifts.

Conclusions:

• CAT provides a straightforward yet highly effective solution for SSDG tasks.
• The method successfully overcomes reliance on fixed thresholds and sensitivity to noisy pseudo-labels.
• Achieves robust generalization in label-efficient settings, enhancing practical DG applicability.