Title:DiCoTTA: Domain-invariant Learning for Continual Test-time Adaptation

Abstract:This paper studies continual test-time adaptation (CTTA), the task of adapting a model to constantly changing unseen domains in testing while preserving previously learned knowledge. Existing CTTA methods mostly focus on adaptation to the current test domain only, overlooking generalization to arbitrary test domains a model may face in the future. To tackle this limitation, we present a novel online domain-invariant learning framework for CTTA, dubbed DiCoTTA. DiCoTTA aims to learn feature representation to be invariant to both current and previous test domains on the fly during testing. To this end, we propose a new model architecture and a test-time adaptation strategy dedicated to learning domain-invariant features without corrupting semantic contents, along with a new data structure and optimization algorithm for effectively managing information from previous test domains. DiCoTTA achieved state-of-the-art performance on four public CTTA benchmarks. Moreover, it showed superior generalization to unseen test domains.