Federated Multi-source Domain Adaptive Object Detection with Probabilistic Teacher
Peggy Joy Lu, Chia-Yung Jui, Jen‐Hui Chuang · 2023
In order to reduce data collecting effort as well as increase data diversity in advanced surveillance systems, collaboration among multiple cameras is practical and efficient. However, training a CNN model at a centralized server may violate user privacy and data confidentiality. We propose a novel scenario wherein data providers (clients) can collaboratively train a model without revealing their datasets and map our problem to multi-source domain adaptive object detection under the federated setting. A Mutual Learning Teacher-Student Framework is setup on clients for domain adaptation. To reduce the false pseudo labels, the technique of Probability Teacher is adopted with the weak-strong augmentation for source/target data. Furthermore, the server fuses models from different clients by model aggregation algorithms, so as to generate a global model for the target domain. Experiments on real world dataset show that our approach outperforms other privacy preserving model aggregation methods.