Exploring wider style space and style shifting based on entropy for domain generalization

Abstract

Domain Generalization (DG) is a research field that enables the neural networks to have generalization capability on the target domain which has not been observed during training. DG approaches are mainly divided into three categories. Among them, a style augmentation method performs well with a relatively simple algorithm. MixStyle and DSU are representative algorithms for solving DG issue. However, it also has limitations, one of which is that a new style generated by these algorithms only occurs in or near the source domain styles during training. This may not improve the generalization capability when the styles of target domains are still not explored by the styles of source and newly generated domains in the entire style space. In this paper, we propose new algorithms for solving the above limitation during training and testing stages. While preserving an original style of source domain, we explore wider style by generating new style via style augmentation. This makes each training sample to explore various styles during training. In the testing stage, we also propose entropy-based style shifting strategy. During testing, our scheme shifts the style of the target domain to the styles of source domains, then passes through the neural network and derives an output from each style. Among them, the one with the minimum entropy of the output is determined as the final output. Experimental results show that our proposed method outperforms existing methods on both multi-source and single-source domain generalization scenarios.