MixStyle-Based Contrastive Test-Time Adaptation: Pathway to Domain Generalization

Kota Yamashita, Kazuhiro Hotta; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 1029-1037

Recent advancements in domain generalization have increasingly focused on Test-time Adaptation (TTA) which adapts models to unknown domains during testing. Test-time Training (TTT) represents a prominent TTA approach utilizing multi-task learning on training images by combining the main task with self-supervised tasks such as rotation prediction and adapting the model to the test domain using only self-supervised tasks during testing. However the selection of appropriate self-supervised tasks poses a challenge in TTT as incorrect choices can degrade model performance. Common self-supervised tasks like rotation prediction are not specifically designed for domain generalization. TENT implements an unsupervised TTA technique utilizing entropy minimization without engaging in self-supervised tasks. Although it bypasses the need for self-supervised tasks its performance can fall short of TTT in certain domains. To address TTT's challenges we propose MixStyle-based Contrastive Test-time Adaptation (MCTTA) which employs the original method of MixStyle-based Contrastive Learning (MCL) to train feature extractors capable of extracting consistent features across different domains. The learning process is divided into Training and TTA phases. During the Training phase the model is generalized to various domains through multi-task learning: main classification task and MCL. In the TTA phase MCL is applied to the test data to adapt the feature extractor to the test domain. By experiments on the DomainBed benchmark library and three datasets (PACS Office-Home and Colored MNIST) MCTTA achieved the highest domain generalization accuracy surpassing not only TTT but also other TTA methods and domain generalization methods.