Continual Test-Time Adaptation With Weighted Contrastive Learning and Pseudo-Label Correction

Real-time adaptability is often required to maintain system accuracy in scenarios involving domain shifts caused by constantly changing environments. While continual test-time adaptation has been proposed to handle such scenarios, existing methods rely on high-accuracy pseudo-labels. Moreover, contrastive learning methods for continuous test-time adaptation consider the aggregation of features from the same class while neglecting the problem of aggregating similar features within the same class. Therefore, we propose “Weighted Contrastive Learning” and apply it to both pre-training and continual test-time adaptation. To address the issue of catastrophic forgetting caused by continual adaptation, previous studies have employed source-domain knowledge to stochastically recover the target-domain model. However, significant domain shifts may cause the source-domain knowledge to behave as noise, thus impacting the model's adaptability. Therefore, we propose “Domain-aware Pseudo-label Correction” to mitigate catastrophic forgetting and error accumulation without accessing the original source-domain data while minimizing the impact on model adaptability. The thorough evaluations in our experiments have demonstrated the effectiveness of our proposed approach.