Domain Generalization

Abstract

Domain Generalization (DG) aims to train a model on one or many source datasets such that it generalizes robustly to a novel target dataset at testing-time, as illustrated in Figure 7.1. Unlike domain adaptation, the target data is not assumed to be available during training in either labeled or unlabeled form. The lack of any available target data makes it more challenging than DA, which typically provides an upper bound for the performance of DG methods. However, the DG problem setting arises widely in practice across diverse areas of computer vision, from earth observation imaging [Koh et al., 2020] to medical imaging [Liu et al., 2020a] and person re-identification [Song et al., 2019]. It is also of academic interest as a type of generalization that humans find effortless but where contemporary computer vision performs poorly [Li et al., 2017a]. DG has therefore been studied increasingly intensively in the last decade [Wang et al., 2021, Zhou et al., 2021a]. In this chapter we highlight the main trends and successes in DG research.