DBPT：基于双存储的连续测试时间自适应提示微调
DBPT:Dual-Bank Prompt Tuning for Continual Test-Time Adaptation
张敏 1吴振宇 1纪阳1
作者信息
- 1. 北京邮电大学信息与通信工程学院,北京 100876
- 折叠
摘要
连续测试时间域自适应（Continual Test-Time Adaptation, CTTA）旨在使预训练模型能够持续适应不断变化的测试域。动态变化的场景通常会导致错误积累和灾难性遗忘,现有的CTTA方法主要依赖单一的知识存储机制,难以同时利用域级和类别级的语义信息。针对这一问题,本文提出一种基于双Bank存储策略的两阶段Prompt自适应方法,通过同时建模域级与类级特异性特征,使冻结的预训练模型在原始特征空间上面对动态场景能够获得更稳健的分类能力。该方法维护了两个Bank：域Bank与类Bank,分别存储历史域的统计信息与各类别的高质量样本原型,并设计两阶段更新策略：阶段一用域Bank生成加权Prompt并进行域级优化,将目标域拉近源域空间；阶段二以第一阶段的Promt为起点实施类级细粒度更新并结合源域原型约束进行类别对齐。实验结果表明,本文方法在多种基准场景下均显著优于现有CTTA方法。
Abstract
Continual Test-Time Adaptation (CTTA) aims to enable pre-trained models to continuously adapt to evolving test domains without access to source data. However, dynamic environments often lead to error accumulation and catastrophic forgetting during online adaptation. Existing CTTA methods mainly rely on a single knowledge storage mechanism, which limits the ability to simultaneously exploit domain-level and class-level semantic information.To address this limitation, a two-stage prompt-based adaptation framework with a dual-bank memory strategy is introduced, explicitly modeling both domain-specific and class-specific characteristics. By operating on the original frozen feature space of pre-trained models, the proposed approach enhances robustness under continuously changing test conditions. Two complementary memory banks are maintained: a Domain Bank, which stores historical domain statistics, and a Class Bank, which preserves high-quality class prototypes. Based on these banks, a two-stage update strategy is designed. In the first stage, weighted prompts are generated by leveraging the Domain Bank to perform domain-level adaptation, aligning the target domain toward the source feature distribution. In the second stage, starting from the adapted prompts obtained in the first stage, fine-grained class-level prompt optimization is conducted under source-domain prototype constraints to achieve precise category alignment.Extensive experiments on multiple CTTA benchmarks demonstrate that the proposed method consistently outperforms existing state-of-the-art CTTA approaches, validating its effectiveness and robustness in dynamic test-time adaptation scenarios.关键词
连续测试时间域自适应/Vision Transformer/提示学习/双Bank存储/两阶段更新Key words
Continual Test-Time Adaptation/ Vision Transformer/ Prompt Learning/ Dual-Bank Storage引用本文复制引用
张敏,吴振宇,纪阳.DBPT：基于双存储的连续测试时间自适应提示微调[EB/OL].(2026-03-02)[2026-03-03].http://www.paper.edu.cn/releasepaper/content/202603-22.学科分类
计算技术、计算机技术
评论