Generalizing to unseen domains via PatchMix

Domain generalization (DG) aims to transfer knowledge learned from multiple source domains to unseen domains. One of the primary challenges hinders DG is the insufficient diversity of source domains, which hampers the model's ability to learn to generalize. Traditional data augmentation methods, which fuse content, style, labels, etc., unable to effectively learn the global features from the source domains. In this paper, we present an innovative approach to domain generalization learning technique, called PatchMix, by stitching the patches of different source domains together to build domain-mixup samples. This approach helps the model to learn the common features of different source domains. Meanwhile, a domain discriminator is introduced to preserve the model's ability to distinguish the source domains, which is proved to be helpful for the model to generalize to unseen domains. To our best knowledge, we are the first to unveil the equation that elucidates the correlation between the number of patches and the number of source domains. Our method, PatchMix, outperforms the current state-of-the-art (SOTA) on four benchmark datasets.