Domain Adversarial SegFormer

Unsupervised domain adaptation (UDA) is a pivotal challenge in machine learning, aiming to transfer knowledge from a labeled source domain to an unlabeled target domain. Despite significant advancements, many UDA methods struggle with real-world applications due to domain discrepancies and the lack of labeled data in the target domain. In this paper, we propose a novel UDA framework that leverages domain-adversarial training and efficient semantic segmentation models to bridge this gap. Our approach involves two key innovations. First, we introduce a domain-adversarial training method specifically designed for segmentation transformers, uncovering new complexities and flexibilities inherent in transformer architectures. Second, we develop an efficient semantic segmentation model tailored for scenarios with limited computational resources and training data, employing an auto-labeling-like Pseudo-Teacher mechanism. Extensive experiments demonstrate that our model achieves a mean Intersection over Union (mIoU) of 61.18% on the challenging GTA → CS, representing 91.92% of oracle performance, while significantly reducing computational requirements. These results highlight the potential of our framework to improve the performance in UDA, making it a viable solution for practical applications.