Abstract: Recent work has shown that deep vision models tend to be overly dependent on low-level or “texture” features, leading to poor generalization. Various data augmentation strategies have been ...