FedS2R paper proposes one-shot federated synthetic-to-real driving segmentation

Researchers propose FedS2R, described on arXiv as the first one-shot federated domain generalization framework for synthetic-to-real semantic segmentation in autonomous driving. The abstract says federated domain generalization has progressed in image classification by enabling collaborative training across clients without sharing raw data, but its use in autonomous-driving semantic segmentation remains underexplored. FedS2R has two components: an inconsistency-driven data augmentation strategy that generates images for unstable classes, and a multi-client knowledge distillation scheme with feature fusion that distills a global model from multiple client models. Experiments on five real-world datasets, Cityscapes, BDD100K, Mapillary, IDD, and ACDC, show the global model significantly outperforms individual client models and trails by only 2 mIoU points a model trained with simultaneous access to all client data, the authors report. They present the results as evidence that FedS2R is effective for synthetic-to-real semantic segmentation under federated learning.