Audio-visual continual test-time adaptation without forgetting

Proposes a method for audio-visual continual test-time adaptation where a source model adapts to unlabeled non-stationary domains with distributional shifts in one or both modalities. Approach handles forgetting during online adaptation.