Entropy Balancing Strategy for Domain Adaptive Semantic Segmentation

In the context of unsupervised domain adaptation for semantic segmentation tasks using self-training methods, pseudo-labels tend to include noise as a result of domain shift and the absence of manual annotations. To address this issue, existing approaches commonly apply entropy regularization to mitigate the noise within pseudo-labels. However, excessive regularization can undermine the reliability of accurate pseudo-labels and cap the effectiveness of the self-training entropy minimization principle. Building upon the integration of source and target domains, this paper introduces an entropy balancing strategy designed to “level the highs and lows” of entropy in the combined domain, thereby placing it in a more balanced state. This approach is advantageous for enhancing prediction values with lower certainty while tempering overconfident predictions. Our method leverages the widely recognized Transformer baseline model, DAFormer, and achieves notable performance gains by incorporating an entropy balance loss into three adaptive semantic segmentation tasks: Cityscapes to ACDC, Cityscapes to FoggyCityscapes + RainCityscapes and Cityscapes to Dark Zurich. Notably, our method delivers competitive outcomes even under conditions of low resolution and constrained graphics memory. The code and data are available at https://github.com/vron8632/EB_DA.