DINO-DIS: Vision Foundation Model-Guided Feature Disentanglement for Adverse Weather Adaptation

Existing unsupervised domain adaptation (UDA) methods rely on labeled source data and unlabeled target data. However, in source-free unsupervised domain adaptation (SFUDA) scenarios, only a source-pretrained model and unlabeled target data are available, posing significant challenges for complex cross-domain alignment tasks such as semantic segmentation from normal to adverse conditions. This paper proposes the scene content consistency hypothesis: target domain images (e.g., rainy, snowy, or low-light scenes) and their corresponding reference images (captured under normal conditions) share prior knowledge of identical scene content. They exhibit high consistency in spatial layout, semantic structure, and static object relationships, while domain differences are confined to low-level perturbations (e.g., illumination, noise, resolution). Based on this hypothesis, we propose DINO-DIS, a method that first designs a feature disentanglement network leveraging the geographic alignment between reference and target images to explicitly decompose backbone features into domain-invariant and domain-specific features. Domain-invariant features then refine pseudo-labels to reduce training noise and enhance the baseline model, enabling cross-domain alignment from normal-condition source models to adverse target domains (e.g., rain, fog, night). Experiments on two normal-to-adverse adaptation benchmarks demonstrate superior semantic segmentation performance.