Abstract
Domain Generalization (DG) is essentially a sub-branch of out-of-distribution generalization, which trains models from multiple source domains and generalizes to unseen target domains. Recently, some domain generalization algorithms have emerged, but most of them were designed with non-transferable complex architecture. Additionally, contrastive learning has become a promising solution for simplicity and efficiency in DG. However, existing contrastive learning neglected domain shifts that caused severe model confusions. In this paper, we propose a Dual-Contrastive Learning (DCL) module on feature and prototype contrast. Moreover, we design a novel Causal Fusion Attention (CFA) module to fuse diverse views of a single image to attain prototype. Furthermore, we introduce a Similarity-based Hard-pair Mining (SHM) strategy to leverage information on diversity shift. Extensive experiments show that our method outperforms state-of-the-art algorithms on three DG datasets. The proposed algorithm can also serve as a plug-and-play module without usage of domain labels.
Abstract (translated)
域扩展(DG)是分布外扩展分支的一个子领域,该方法从多个源域中训练模型,并扩展到未知的目标域。最近,一些域扩展算法已经出现,但大多数设计使用了不可转移的复杂架构。此外,对比学习已成为在域扩展方面的一个有前途的解决方案,以简化和效率为关键。然而,现有的对比学习忽视了导致严重模型混淆的域转换。在本文中,我们提出了基于特征和原型对比的二元对比学习(DCL)模块,并设计了一个独特的因果融合注意力(CFA)模块,将单个图像的多个视角融合成一个原型,同时引入基于相似性的 Hard-pair Mining(SHM)策略,以利用多样性转移的信息。广泛的实验结果表明,我们的方法在三个域扩展数据集上优于当前最先进的算法。该提议算法还可以用作可插拔模块,而不需要域标签。
URL
https://arxiv.org/abs/2301.09120