Style Balancing and Test-Time Style Shifting for Domain Generalization

Jungwuk Park, Dong-Jun Han, Soyeong Kim, Jaekyun Moon

Keywords: Domain generalization, Style mixing, Arbitrary style transfer

TL;DR: We propose style balancing and test-time style shifting for domain generalization, to handle the imbalance issues and the issue on large style gap between source and target domains.

Abstract:

Given a training set that consists of multiple source domains, the goal of domain generalization (DG) is to train the model to have generalization capability on the unseen target domain. Although various solutions have been proposed, existing ideas suffer from severe cross-domain data/class imbalance issues that naturally arise in DG. Moreover, the performance of prior works are degraded in practice
where the gap between the style statistics of source and target domains is large. In this paper, we propose a new strategy to handle these issues in DG. We first propose style balancing, which strategically balances the number of samples for each class across all source domains in the style-space, providing a great platform for the model to get exposed to various styles per classes during training. Based on
the model trained with our style balancing, we also propose test-time style shifting, which shifts the style of the test sample (that has a large style gap with the source domains) to the nearest source domain that the model is already familiar with, to further improve the prediction performance. Our style balancing and test-time style shifting work in a highly complementary fashion, and can successfully work in conjunction with various other DG schemes. Experimental results on benchmark datasets show the improved performance of our scheme over existing methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning