Domain generalization via Inter-domain Alignment and Intra-domain Expansion

The performance of traditional deep learning models tends to drop dramatically during being deployed in real-world scenarios when the distribution shift between the seen training and unseen test data occurs. Domain Generalization methods are designed to achieve generalizability to deal with the above issue. Since the features extracted by softmax cross-entropy loss are not adequately domain-invariant, previous works in Domain Generalization have attempted to overcome this problem by employing contrastive-based losses which pull positive pairs (i.e., samples with the same class label) from different domains closer. Unfortunately, these approaches tend to produce an extremely small feature space, which is not robust facing unseen domain and easily overfits to source domains. To address the aforementioned issue, we propose a novel loss named IAIE Loss to simultaneously perform Inter-domain Alignment and Intra-domain Expansion for positive pairs, which facilitates the model to extract domain-invariant features and mitigates overfitting. Specifically, we design two sets of positive samples named "easy positive samples"and "hard positive samples". IAIE Loss pulls the hard positive pairs closer (alignment) while pushing the easy positive pairs apart (expansion). The state-of-the-art results on multiple DG benchmark datasets verify the effectiveness of our method.