DWDM: Dynamically Weighted Three-Domain Mixing for Domain-Adaptive Semantic Segmentation

In unsupervised domain adaptation (UDA), the aim is to adapt a model trained on source data to target data without having access to target annotations. However, when the dissimilarity between the source and target domains is substantial, previous UDA methods for semantic segmentation often struggle to achieve satisfactory segmentation performance. To address this issue, we propose dynamically weighted three-domain mixing (DWDM) to address the challenge of poor segmentation performance in UDA task when the source and target domains differ significantly. DWDM achieves good segmentation performance in the target domain by using dynamically weighted three-domain mixing and a dual-teacher network. Norms are also added to the loss function to ensure smoother network performance improvement. DWDM significantly improves semantic segmentation tasks in clear to adverse weather UDA scenarios. Experimental results demonstrate an unprecedented UDA performance of \(70.1\%\) mIoU on Cityscapes to ACDC, an improvement of \(2.1\%\) compared to the previous state-of-the-art.