Abstract
Unsupervised domain adaptation for semantic segmentation aims to transfer knowledge from a labeled source domain to another unlabeled target domain. However, due to the label noise and domain mismatch, learning directly from source domain data tends to have poor performance. Though adversarial learning methods strive to reduce domain discrepancies by aligning feature distributions, traditional methods suffer from the training imbalance and feature distortion problems. Besides, due to the absence of target domain labels, the classifier is blind to features from the target domain during training. Consequently, the final classifier overfits the source domain features and usually fails to predict the structured outputs of the target domain. To alleviate these problems, we focus on enhancing the adversarial learning based feature alignment from three perspectives. First, a classification constrained discriminator is proposed to balance the adversarial training and alleviate the feature distortion problem. Next, to alleviate the classifier overfitting problem, self-training is collaboratively used to learn a domain robust classifier with target domain pseudo labels. Moreover, an efficient class centroid calculation module is proposed and the domain discrepancy is further reduced by aligning the feature centroids of the same class from different domains. Experimental evaluations on GTA5 <inline-formula><tex-math notation="LaTeX">\rightarrow</tex-math></inline-formula> Cityscapes and SYNTHIA <inline-formula><tex-math notation="LaTeX">\rightarrow</tex-math></inline-formula> Cityscapes demonstrate state-of-the-art results compared to other counterpart methods. The source code and models have been made available at.<xref ref-type="fn" rid="fn1"> 1 1
[Online]. Available: https://github.com/NUST-Machine-Intelligence-Laboratory/EFA .