Practical Single Domain Generalization via Training-time and Test-time Learning

Single domain generalization aims to learn a model that generalizes well to unseen target domains by using a related source domain. However, most existing methods only focus on improving the generalization performance of the model during training, making it difficult to achieve satisfactory performance when deployed in the target domain with large domain shifts. In this paper, we propose a Practical Single Domain Generalization (PSDG) method, which first leverages the knowledge in a source domain to establish a model with good generalization ability in the training phase, and subsequently updates the model to adapt to target domain data using knowledge in the unlabeled target domain during the testing phase. Specifically, during training, PSDG leverages a newly proposed style (e.g., background features) generator named StyIN to generate novel domain data. Moreover, PSDG introduces style-diversity regularization to constantly synthesize distinct styles to expand the coverage of training data, and introduces object-consistency regularization to capture consistency between the currently generated data and the original data, making the model filter style knowledge during training. During testing, PSDG uses a sample-aware and sharpness-aware minimization method to seek for a flat entropy minimum surface for further model optimization by using the knowledge in the unlabeled target domain. Using three real-world datasets the experiments have demonstrated the effectiveness of PSDG, in comparison with several state-of-the-art methods.