FedS2R: One-Shot Federated Domain Generalization for Synthetic-to-Real Semantic Segmentation in Autonomous Driving
Researchers have introduced FedS2R, a novel one-shot federated domain generalization framework specifically designed for synthetic-to-real semantic segmentation in autonomous driving. This framework addresses the challenge of training models across different datasets without sharing raw data, a crucial aspect for privacy and efficiency. FedS2R incorporates an inconsistency-driven data augmentation strategy and a multi-client knowledge distillation scheme with feature fusion to improve model performance. Experiments on multiple real-world datasets demonstrate that FedS2R significantly enhances model capabilities compared to individual client models, achieving results close to models trained with full data access. AI
IMPACT Enhances synthetic-to-real data transfer for autonomous driving systems, potentially improving model robustness and reducing reliance on extensive real-world data collection.