[Re]Domain Generalization with MixStyle

Anonymous

Keywords: domain generalization, classification

Abstract: To strengthen the generalization ability of CNNs, diverse source data from multiple relevant heterogeneous domains is collected so that CNN model is allowed to learn more domain-invariant features, and hence generalize to unseen out-of-distribution data. This problem is largely studied under domain generalization(DG). Among recent DG methods, a widely used assumption is that images are generated from style and semantic information or disentangled into style features and semantic features, whose semantic information is domain-agnostic while the style information is domain-specific. MixStyle is proposed to mix the domain-related features directly after residual block in a similar way borrowed from AdaIN. In original paper, MixStyle appears to work in three totally different tasks including category classification, instance retrieval and reinforcement learning. For the sake of research interest and knowledge limitation, we reproduce the classification task on PACS dataset with sufficient experiments to exhibit how MixStyle works in CNNs and improves performance of classification. Furthermore, we are not only re-implementing MixStyle, but also extend MixStyle to MixAll, where mixed statistic features across all source domains would perform more stable.

Paper Url: https://openreview.net/forum?id=6xHJ37MVxxp