Nam Hyeonseob ; Lee HyunJae ; Park Jongchan ; Yoon Wonjun ; Yoo Donggeun

Abstract

Convolutional Neural Networks (CNNs) often fail to maintain their performancewhen they confront new test domains, which is known as the problem of domainshift. Recent studies suggest that one of the main causes of this problem isCNNs' strong inductive bias towards image styles (i.e. textures) which aresensitive to domain changes, rather than contents (i.e. shapes). Inspired bythis, we propose to reduce the intrinsic style bias of CNNs to close the gapbetween domains. Our Style-Agnostic Networks (SagNets) disentangle styleencodings from class categories to prevent style biased predictions and focusmore on the contents. Extensive experiments show that our method effectivelyreduces the style bias and makes the model more robust under domain shift. Itachieves remarkable performance improvements in a wide range of cross-domaintasks including domain generalization, unsupervised domain adaptation, andsemi-supervised domain adaptation on multiple datasets.