Learning Robust Shape-Based Features for Domain Generalization

Domain generalization is a challenging problem of learning models that can generalize to novel testing domains which are unavailable during training and follow different distributions from training domains. In this paper, we introduce a simple but effective method for domain generalization, which is based on the object shape hypothesis. The main idea is from some recent studies which demonstrate that leading the network to focus more on the object shape and recognizing objects through shape-based features are more significant and robust. To achieve this, we first stylize source domain images into randomly sampled styles through a style transfer network and then use both the stylized images and original source domain images to learn the classification model. With images of the same content but different styles (textures), we expect the network to classify according to the global object shape rather than local textures and further transfer the object shape knowledge to novel domains. By this way, the generalization ability of the classification model will be improved. We conduct extensive experiments on PACS, VLCS and OfficeHome benchmarks and both the quantitative and qualitative analysis demonstrate the effectiveness of the proposed method.