Abstract
Continual Test-Time Adaptation (CTTA) seeks to adapt source pre-trainedmodels to continually changing, unseen target domains. While existing CTTAmethods assume structured domain changes with uniform durations, real-worldenvironments often exhibit dynamic patterns where domains recur with varyingfrequencies and durations. Current approaches, which adapt the same parametersacross different domains, struggle in such dynamic conditions-they faceconvergence issues with brief domain exposures, risk forgetting previouslylearned knowledge, or misapplying it to irrelevant domains. To remedy this, wepropose DPCore, a method designed for robust performance across diverse domainchange patterns while ensuring computational efficiency. DPCore integratesthree key components: Visual Prompt Adaptation for efficient domain alignment,a Prompt Coreset for knowledge preservation, and a Dynamic Update mechanismthat intelligently adjusts existing prompts for similar domains while creatingnew ones for substantially different domains. Extensive experiments on fourbenchmarks demonstrate that DPCore consistently outperforms various CTTAmethods, achieving state-of-the-art performance in both structured and dynamicsettings while reducing trainable parameters by 99% and computation time by 64%compared to previous approaches.