CLASH-CTTA: Class-Wise Shift-Aware Hierarchical Continual Test-Time Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Domain shift between training and test data hinders deep model generalization.
• Continual Test-Time Adaptation (CTTA) addresses evolving target distributions using online test data.
• Existing CTTA methods often isolate batch processing or reliable samples, limiting combined advantages.

Purpose of the Study:

• To propose a fully source-free CTTA method addressing limitations of existing approaches.
• Introduce CLAss-wise Shift-aware Hierarchical Continual Test-Time Adaptation (CLASH-CTTA) for enhanced domain adaptation.

Main Methods:

• Employs a hierarchical updating strategy combining slow learning of general representations and fast learning of domain-specific knowledge.
• Maintains a Class-wise Shift-aware Representative Set to mitigate discrepancies in class sensitivity to domain shifts.
• Utilizes Spearman's rank correlation for sample filtering and alignment to class prototypes.

Main Results:

• CLASH-CTTA demonstrates superior performance compared to state-of-the-art methods.
• Effectiveness validated across corruption and natural domain shift datasets.
• Achieves strong results in continual, gradual, and diverse batch size scenarios.

Conclusions:

• CLASH-CTTA offers a robust and effective solution for continual test-time adaptation.
• The proposed hierarchical learning and class-wise representative strategies significantly improve deep model generalization under domain shift.