D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Frequency and Pixel Spaces
D-GAP is a method to boost AI model robustness when applied to new environments (OOD). It introduces datasetagnostic augmentations by targeting both frequency and pixel spaces using gradient information. This dual approach allows the model to handle domain shifts better than standard methods, leading to significant accuracy gains on real-world and benchmark datasets.