Source-Free Controlled Adaptation of Teachers for Continual Test-Time Adaptation

TL;DR

The paper introduces a source-free continual test-time adaptation method that dynamically adjusts the momentum in a mean teacher framework based on incoming data quality, reducing model drift. It uses class prototypes from a source model to align target data without requiring source data access. The approach achieves state-of-the-art performance on benchmark datasets.

Why It Matters

Real-world deployment often faces evolving data distributions, and this method enables robust adaptation without relying on source data, enhancing practicality and security in production systems. It improves model reliability over time while maintaining privacy and reducing dependency on training data.

Key Findings

• Dynamic momentum adjustment in the mean teacher reduces drift during continual adaptation.
• Class prototype estimation from the source model enables effective alignment of target data without source data access.
• The method outperforms existing state-of-the-art approaches on benchmark datasets, including those requiring source data.