Data Distribution Transfer for Out Of Distribution Generalization

Modern deep neural networks suffer from performance degradation when evaluated on testing data under different distributions from training data. The goal of out-of-distribution generalization is to solve this problem by learning transferable knowledge from source domains to generalize to invisible target domains. This paper presents a data augmentation method for out-of-distribution generalization. The main assumption is that the main data distribution of an image mostly contains domain-related information, such as color, illumination, texture content, etc, which hurts the domain shifts. To force the model to pay less attention to this part of the information, we propose a new data augmentation method based on the main distribution transition. Extensive experiments on two data set have demonstrated that the proposed method is able to achieve state-of-the-art performance for domain generalization. At the same time, our method can not only combine with other methods to produce a superposition generalization effect but also generate obfuscation data cheaply.