Unsupervised Domain Adaptation for Semantic Segmentation with Global and Local Consistency

Xiangxuan Shan, Zijin Yin, Jiayi Gao, Kongming Liang, Zhanyu Ma, Jun Guo

Abstract: Unsupervised domain adaptation(UDA) for semantic segmentation aims to learn from labeled synthetic data to segment the unlabeled real data. Many recent methods use generative networks to acquire real-like images for mitigating domain shift. However, these methods only ensure global style consistency between two domains and fail to impose pixel-wise constraint which is also referred to as local content consistency. To address the above problem, we propose a global and local consistency network to reduce the domain gap in unsupervised domain adaptation for semantic segmentation. To this end, we first constrain global style consistency through a generative adversarial network to acquire real-like latent domain images. Then we enhance local content consistency based on pixel-wise entropy minimization. Experimental results show that our method has superiority over other competitive methods on GTA5 $$\rightarrow $$ Cityscapes.