Reducing Domain Gap by Reducing Style Bias

TL;DR

This paper introduces SagNets, a method that reduces CNNs' style bias by disentangling style from content, significantly improving robustness across various cross-domain tasks.

Contribution

The paper proposes a novel style-agnostic network architecture that disentangles style and content to mitigate domain shift in CNNs.

Findings

SagNets reduce style bias effectively.

Improved performance in domain generalization and adaptation.

Robustness across multiple datasets and tasks.

Abstract

Convolutional Neural Networks (CNNs) often fail to maintain their performance when they confront new test domains, which is known as the problem of domain shift. Recent studies suggest that one of the main causes of this problem is CNNs' strong inductive bias towards image styles (i.e. textures) which are sensitive to domain changes, rather than contents (i.e. shapes). Inspired by this, we propose to reduce the intrinsic style bias of CNNs to close the gap between domains. Our Style-Agnostic Networks (SagNets) disentangle style encodings from class categories to prevent style biased predictions and focus more on the contents. Extensive experiments show that our method effectively reduces the style bias and makes the model more robust under domain shift. It achieves remarkable performance improvements in a wide range of cross-domain tasks including domain generalization, unsupervised domain adaptation, and semi-supervised domain adaptation on multiple datasets.