FedDAD: Federated Domain Adaptation for Object Detection

Training an object detection model often requires numerous annotated images on a centralized host, which may violate user privacy and data confidentiality. Federated learning (FL) resolves this issue by allowing multiple clients, e.g., cameras, to collaboratively train while protecting privacy. However, models trained with FL fail be generalized for new target domain due shift when the between source domains are statistically different. In work, we formulate real-world problem as source-free multi-domain adaptation in architecture. Moreover, propose adaptive algorithm, called FedDAD (Federated Domain Adaptive Detector), aggregates dynamic attention targeting unsupervised server, utilize instance-level alignment alleviate effects of scene variation clients. Experimental results show that improves average precision (AP) up 10.05% 19.15% compared popular FedAvg specific classes KAIST MI3 datasets, respectively.