Enhancing Semantic Segmentation in Open Compound Domain Adaptation Through Mixed Image and Epistemic Uncertainty

Open Compound Domain Adaptation (OCDA) presents a novel challenge in semantic segmentation, where the target domain combines multiple domains with blurry boundaries and unseen categories. While UDA-based semantic segmentation achieves high accuracy on unseen domain data, it struggles to maintain accuracy in open compound domains. Specifically, data augmentation for accurate predictions is challenging, and uncertainty in prediction probabilities often goes overlooked when encountering unknown categories from new domains. In this paper, we propose an uncertainty quantification method to measure the epistemic uncertainty of the model, thereby improving the reliability of its generated predictions. We also propose a novel data augmentation approach that combines paired images from different domains, employing Global Luminous Alignment (GLA) to generate new augmented samples, thereby reducing the domain variance between the target and source domain data. Experiments on GTA5, BDD100K, Synthia, and Cityscapes datasets demonstrate the effectiveness of our methods.