Improving Domain Generalization with Interpolation Robustness

Domain generalization (DG) uses multiple source (training) domains to learn a model that generalizes well to unseen domains. Existing approaches to DG need more scrutiny over (i) the ability to imagine data beyond the source domains and (ii) the ability to cope with the scarcity of training data. To address these shortcomings, we propose a novel framework - \emph{interpolation robustness}, where we view each training domain as a point on a domain manifold and learn class-specific representations that are domain invariant across all interpolations between domains. We use this representation to propose a generic domain generalization approach that can be seamlessly combined with many state-of-the-art DG methods. Through extensive experiments, we show that our approach can enhance the performance of several methods, namely, DeepAll, DIRT and DGER in the conventional and the limited training data setting across multiple datasets.