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 a model while protecting user...
Saved in:
Published in | IEEE access Vol. 11; pp. 51320 - 51330 |
---|---|
Main Authors | , , |
Format | Journal Article |
Language | English |
Published |
Piscataway
IEEE
2023
The Institute of Electrical and Electronics Engineers, Inc. (IEEE) |
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | 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 a model while protecting user privacy. However, models trained with FL may fail to be generalized for new target domain due to domain shift when the data between source and target domains are statistically different. In this work, we formulate a real-world object detection problem as a source-free multi-domain adaptation problem in FL architecture. Moreover, we propose an adaptive FL algorithm, called FedDAD (Federated Domain Adaptive Detector), which aggregates models with dynamic attention targeting the unsupervised domain on server, and utilize instance-level alignment to alleviate the effects of scene variation on clients. Experimental results show that FedDAD improves the average precision (AP) by up to 10.05% and 19.15% compared to the popular FedAvg for specific object classes in the KAIST and MI3 datasets, respectively. |
---|---|
ISSN: | 2169-3536 2169-3536 |
DOI: | 10.1109/ACCESS.2023.3279132 |