上海科技大学知识管理系统

Test-Time Adaptation (TTA) requires adapting a source-domain model to the target domain using online test data inputs. However, the issue of error accumulation has become one of the major obstacles limiting the model's long-term adaptation performance. Many existing methods attempt to mitigate error accumulation by filtering out noisy samples based on a single criterion, which often results in insufficient adaptation to the target domain. Moreover, the problem of imbalanced predicted label distribution is overlooked, leading to unbalanced adaptation. To address these issues, we propose a Reliable and Balanced test-time adaptation method (ReBa) based on multiple loss weighting mechanism. ReBa incorporates two types of loss weighting criteria: (1) Historical Similarity Weighting, which evaluates the similarity between the current prediction and the historical average prediction, encouraging the model to make more diverse predictions. (2) Class Prediction Frequency Weighting, which assigns higher weights to predictions of low-frequency classes while integrating entropy-based weighting, encouraging balanced and reliable prediction. By leveraging multiple loss weighting mechanism, ReBa enables deep, balanced, and reliable adaptation in dynamically changing target domains. Extensive experiments on corruption and natural shift datasets demonstrate the effectiveness of the proposed method.