A Privacy-Preserving Approach for Multi-Source Domain Adaptive Object Detection

A new research topic, multi-source domain adaptive object detection (MSDAOD) under privacy-preserving constraint is explored in this paper, where the clients can only access their own source data while the server can only access unlabeled target data. Accordingly, a novel MSDAOD framework is proposed wherein the clients employ a source-only Probabilistic Faster R-CNN (PFRCNN) to generate models with localization uncertainty, while a Multi-teacher Pseudo-label Ensemble Network (MPEN) is developed on the server side. In MPEN, FedMA-based algorithm aggregates the above models to a domain-invariant backbone while a novel pseudo-label ensemble (PLE) scheme is employed to reduce false positives arising from domain specific parts, and enhance the overall system performance using target domain information. Experiments demonstrate that our method outperforms other state-of-the-art MSDAOD and privacy-preserving methods by 10%∼16% in average precision (AP).