Domain Generalization (DG) studies the capability of a deep learning model to generalize to out-of-training distributions. In the last decade, literature has been massively filled with training methodologies that claim to obtain more abstract and robust data representations to tackle domain shifts. Recent research has provided a reproducible benchmark for DG, pointing out the effectiveness of naive empirical risk minimization (ERM) over existing algorithms. Nevertheless, researchers persist in using the same outdated feature extractors, and no attention has been given to the effects of different backbones yet. In this paper, we start back to the backbones proposing a comprehensive analysis of their intrinsic generalization capabilities, which so far have been ignored by the research community. We evaluate a wide variety of feature extractors, from standard residual solutions to transformer-based architectures, finding an evident linear correlation between large-scale single-domain classification accuracy and DG capability. Our extensive experimentation shows that by adopting competitive backbones in conjunction with effective data augmentation, plain ERM outperforms recent DG solutions and achieves state-of-the-art accuracy. Moreover, our additional qualitative studies reveal that novel backbones give more similar representations to same-class samples, separating different domains in the feature space. This boost in generalization capabilities leaves marginal room for DG algorithms. It suggests a new paradigm for investigating the problem, placing backbones in the spotlight and encouraging the development of consistent algorithms on top of them. The code is available at https://github.com/PIC4SeR/Back-to-Bones.

翻译：暂无翻译