VITA: A Multi-Source Vicinal Transfer Augmentation Method for Out-of-Distribution Generalization
VITA addresses poor robustness to image corruption by generating diverse, high-quality samples that stay close to the original data manifold. It combines tangent transfer, which uses vicinal differences to approximate local structures, with a generative model that integrates multi-source vicinal samples. This process creates on-manifold augmented data superior to existing methods.