Decorate the Newcomers: Visual Domain Prompt for Continual Test Time Adaptation

Continual Test-Time Adaptation (CTTA) aims to adapt the source model continually changing unlabeled target domains without access data. Existing methods mainly focus on model-based adaptation in a self-training manner, such as predicting pseudo labels for new domain datasets. Since are noisy and unreliable, these suffer from catastrophic forgetting error accumulation when dealing with dynamic data distributions. Motivated by prompt learning NLP, this paper, we propose learn an image-layer visual while having parameters frozen. During testing, datasets can be adapted reformulating input learned prompts. Specifically, devise two types of prompts, i.e., domains-specific prompts domains-agnostic extract current knowledge maintain domain-shared continual adaptation. Furthermore, design homeostasis-based strategy suppress domain-sensitive domain-invariant more effectively. This transition model-dependent paradigm model-free one enables us bypass problems. Experiments show that our proposed method achieves significant performance gains over state-of-the-art four widely-used benchmarks, including CIFAR-10C, CIFAR-100C, ImageNet-C, VLCS