Abstract
Domain generalization (DG) and unsupervised domain adaptation (UDA) aim to solve the domain-shift problem that arises when the trained model is tested in the domain with different style distribution from the training data. Style Normalization and Restitution(SNR) has solved this problem to a certain extent and achieved the best performance. However, SNR ignores the discriminative information encoded in the appearance style information, which limits the performance of the model. In this paper, we propose Two-level Style Normalization and Restitution(Tl-SNR) to solve this problem. First, we use group whitening to introduce the appearance style information encoded in the second-order statistic into the SNR, which prepares for restituting the task-relevant discriminative information in the appearance information later. Secondly, we defined dynamic affine parameters, which improves the affine parameters in group whitening. It makes the model adjust adaptively according to the characteristics of the sample, so as to better exploit the capabilities of the model. Finally, we designed a Two-level Style Normalization and Restitution module based on the improved group whitening for domain generalization and unsupervised domain adaptation. Extensive experiments show that our method is effective. And our method outperforms state-of-the-art DG and UDA methods on four benchmarks.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call