Improving Domain Generalization using Style Regularization

We study the problem of improving domain generalization on deep networks by reducing the bias towards texture learned by these models when pre-trained in large color image datasets like ImageNet. To do so, we present a style regularization to enforce more shape-biased learning. Also, we propose an experimental setup using synthetically created test sets using state-of the-art style transfer methods. We report our experiments on stylized versions of CIFAR-10 and STL-10 datasets. In our preliminary results presented here, we show that our style regularization improves performance on new domains but not as significantly as with style augmentation.