Domain Generalization without Source Data via Domain Prompt-Estimator Joint Training

We tackle the source-free domain generalization (SFDG) that aims to train a model generalizable to unseentarget domains without using any actual source data. Existing SFDG methods have pragmatically tackled this challengingproblem by using large-scale vision-language pre-trained models (VLMs), but they do not exploit the fullpotential of VLMs due to the lack of domain flexibility. Specifically, they employ domain-shared configurations(e.g., classifier, representation) for data inputs from any domain, even though these configurations should be adjustedaccording to the differences in various domains. Our method endows the model with domain flexibility by jointlytraining the following two modules: (1) various domain-specific prompts to ensure generalizability to unseen domainsand (2) a domain estimator to select favorable domain-specific prompts for the input image. Notably, thistraining can be realized in the form of simple domain classification learning. Additionally, our method can easilyboost the recognition accuracy further by incorporating a simple yet effective test-time adaptation technique. We empiricallydemonstrate the superiority of our method to state-of-the-art SFDG methods on four domain generalizationbenchmarks.