Abstract:
This paper tackles domain generalization (DG) problem, the task of utilizing only source domain(s) to learn a model that generalizes well to unseen domains. A key challenge faced by DG is often the limited diversity in available source domain(s) that restricts the network's ability in learning a generalized model. Existing DG approaches leveraging data augmentation to address this problem mostly rely on compute-intensive auxiliary networks coupled with various losses and also suffer from additional training overhead. To this end, we propose a simple and efficient DG approach to augment source domain(s). We hypothesize the existence of favourable correlation between the source and target domain's major modes of variation, and upon exploring those modes in the source domain we can realize meaningful alterations to {background, appearance, pose and texture of object classes}. Inspired by this, our new DG approach performs feature-space augmentation by identifying the dominant modes of change in the source domain and implicitly including the augmented versions along those directions to achieve a better generalization across domains. Our method shows competitive performance against the current state-of-the-art methods on three popular DG benchmarks. Further, encouraging results on challenging single-source setting validate strong domain generalization capabilities of our approach.
