Title: FedS2R: Enabling One-Shot Federated Domain Generalization for Synthetic-to-Real Semantic Segmentation in Autonomous Driving

Abstract:

While federated domain generalization has demonstrated significant potential in image classification by facilitating collaborative model training across distributed clients without the exchange of raw data, its application to semantic segmentation in autonomous driving remains largely uncharted. To address this gap, we introduce FedS2R, pioneering the first one-shot federated domain generalization framework tailored for synthetic-to-real semantic segmentation in this domain. The architecture of FedS2R integrates two key mechanisms: a data augmentation strategy driven by inconsistency to synthesize images for classes that exhibit instability, and a multi-client knowledge distillation approach featuring feature fusion to aggregate a global model from various client models. We evaluated the framework using five real-world datasets: Cityscapes, BDD100K, Mapillary, IDD, and ACDC. Our findings indicate that the resulting global model substantially surpasses individual client models, trailing only by 2 mIoU points compared to a model trained with direct access to all client data. These outcomes underscore the efficacy of FedS2R in advancing synthetic-to-real semantic segmentation within autonomous driving systems under a federated learning paradigm.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC