DFDW: Distribution-aware Filter and Dynamic Weight for open-mixed-domain Test-time adaptation

Test-time adaptation (TTA) aims to adapt the pre-trained model to the unlabeled test data stream during inference. However, existing state-of-the-art TTA methods typically achieve superior performance in closed-set scenarios, and often underperform in more challenging open mixed-domain TTA scenarios. This can be attributed to ignoring two uncertainties: domain non-stationarity and semantic shifts, leading to inaccurate estimation of data distribution and unreliable model confidence. To alleviate the aforementioned issue, we propose a universal TTA method based on a Distribution-aware Filter and Dynamic Weight, called DFDW. Specifically, in order to improve the model’s discriminative ability to data distribution, our DFDW first designs a distribution-aware threshold to filter known and unknown samples from the test data, and then separates them based on contrastive learning. Furthermore, to improve the confidence and generalization of the model, we designed a dynamic weight consisting of category-reliable weight and diversity weight. Among them, category-reliable weight uses prior average predictions to enhance the guidance of high-confidence samples, and diversity weight uses negative information entropy to increase the influence of diversity samples. Based on the above approach, the model can accurately identify the distribution of semantic shift samples, and widely adapt to the diversity samples in the non-stationary domain. Extensive experiments on CIFAR and ImageNet-C benchmarks show the superiority of our DFDW.