DPAC: Dual-Path Domain Bridging With Atmospheric Correction for Adverse Weather Adaptation

Data-driven semantic segmentation models often generalize poorly to unseen adverse weather conditions due to significant domain shifts and the lack of labeled data in the target domain. While recent unsupervised domain adaptation (UDA) approaches attempt to mitigate this gap through adversarial alignment or self-training, they typically treat domain discrepancy as a holistic shift, leading to suboptimal adaptation under diverse and complex weather conditions. We propose a Dual-Path Domain Bridging with Atmospheric Correction (DPAC) framework that disentangles style-related and weather-related discrepancies via two complementary teacher models. First, a reference domain is introduced to decompose the source–target gap, and the teacher models are trained using dual-path mixed data to mitigate discrepancies caused by style variation and weather degradation. Then, an atmospheric correction module is integrated into the weather-path teacher to restore degraded target images and improve feature-level alignment. Lastly, a student model is optimized under the alternating supervision of both teachers, enabling robust knowledge integration and enhanced generalization. Extensive experiments on public benchmarks demonstrate the superiority of our approach. DPAC achieves significant performance gains of 5.1% on the Cityscapes to ACDC task, 6.1% on Cityscapes to Foggy Driving, and 5.3% on Cityscapes to Foggy Zurich over a strong baseline, validating the effectiveness of the dual-path domain bridging strategy in complex multi-weather scenarios.