Uncertainty-aware consistency regularization for cross-domain semantic segmentation

Unsupervised domain adaptation (UDA) aims to adapt existing models of the source a new target with only unlabeled data. Most methods suffer from noticeable negative transfer resulting either error-prone discriminator network or unreasonable teacher model. Besides, local regional consistency in UDA has been largely neglected, and extracting global-level pattern information is not powerful enough for feature alignment due abuse use contexts. To this end, we propose an uncertainty-aware regularization method cross-domain semantic segmentation. Firstly, introduce uncertainty-guided loss dynamic weighting scheme by exploiting latent uncertainty samples. As such, more meaningful reliable knowledge model can be transferred student We further reveal reason why current often unstable minimizing discrepancy. design ClassDrop mask generation algorithm produce strong class-wise perturbations. Guided mask, ClassOut strategy realize effective fine-grained manner. Experiments demonstrate that our outperforms state-of-the-art on four benchmarks, i.e., GTAV $\rightarrow $ Cityscapes SYNTHIA Cityscapes, Virtual KITTI $\rightarrow$ KITTI.