Coarse-to-fine online latent representations matching for one-stage domain adaptive semantic segmentation

Zihao Dong, Sijie Niu, Xizhan Gao, Xiuli Shao

Abstract: Highlights•Coarse-to-fine online matching fuses multi-level features for unified domain alignment.•Region-level self-supervised learning rectifies domain shift without extra parameter.•The predicted discrepancy between two classifiers denotes the uncertainty in network.•Experiments show the effectiveness of one-stage domain adaptive segmentation.