Test-time adaptation (TTA) aims to make the trained neural network model adapt to the test data distribution at test time using unlabeled test data. Existing TTA methods mainly consider adaptation in a single or multiple static environments. However, in non-stationary environments, the test data distribution changes continuously over time, which leads to the instability of previous TTA methods. Therefore, we propose a robust and reliable symmetric cross entropy test-time adaptation algorithm (RRSTA) based on robust and reliable symmetric cross entropy. First, in order to improve the robustness to noise distribution changes and alleviate catastrophic forgetting, we propose a symmetric cross entropy based on the mean teacher model, which encourages the model to predict correctly and punishes the wrong prediction. Secondly, in order to improve the robustness to different noise samples, we propose a dual-stream perturbation technique, which guides the weak-to-strong perturbation view of the student model through the strong view of the teacher model. Finally, we propose a reliable entropy minimization strategy to prevent the drastic change of parameters and stabilize adaptation. Extensive experiments and ablation studies on CIFAR10C and CIFAR100C confirm the effectiveness of the proposed method. Compared with the unadapted model, the error rate is significantly reduced by 26.13% and 14.69%, and it is significantly better than the second-best method.

[2024-02-02] Accepted Paper /doc/2023.10.0500/v1