Domain generalization via geometric adaptation over augmented data

Abstract

This article addresses the challenge of adapting deep learning models trained on specific datasets to effectively generalize to similar-class dataset with different underlying distributions. We introduce a novel deep representation learning method that takes into account both statistical and geometric properties of features for domain generalization. Our approach utilizes Fourier augmentation and Nyström estimation to evaluate the similarity between graphs derived from original and augmented data features. Furthermore, we employ a contrastive loss function to maintain proximity among samples belonging to the same class while ensuring separation between samples from different classes in the feature space. By minimizing these loss functions, our method aims to enhance model generalizability across diverse domains. Comprehensive experiments conducted on real-world benchmark datasets, including PACS, Office-Home, VLCS, Digits-DG and UTKFace, demonstrate the effectiveness of the proposed method. The results consistently indicate superior performance compared to other approaches under various conditions, underscoring its robustness in achieving improved generalization across domains.