Reducing Domain Gap by Reducing Style Bias
This addresses domain shift issues for computer vision applications, offering a novel approach to enhance model robustness in cross-domain scenarios.
The paper tackles the problem of domain shift in CNNs by reducing their bias towards image styles, proposing Style-Agnostic Networks (SagNets) that disentangle style encodings to focus on content, resulting in remarkable performance improvements across various cross-domain tasks.
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.