Communication-efficient Multi-source Domain Adaptive Object Detection under Privacy Constraints

To establish a more generalized model in object detection, collaboration among multiple cameras can increase data diversity and reduce the effort for data collection, leading to a new research area as multi-source domain adaptative object detection (MSDAOD). However, preserving source data privacy in MSDAOD is challenging due to the lack of information integrated from all source domains. In this paper, we present an architecture that allows multiple clients protect the privacy of their own local data while the server only access target data. First, we analyze the effectiveness of using multiple sources, in domain adaptive object detection task. In client sides, we propose a source-only probabilistic teacher (PT) and leverage probabilistic teacher for domain adaptation (PTDA) as detectors to reduce false negatives. Moreover, we also introduce a Pseudo-label Voting Mechanism to filter out false positives with minimal communication costs. The performance of the proposed approach is evaluated on the ck2b and skf2c datasets and compared with other multi-source domain adaptation as well as federated learning methods. To sum up, the proposed method achieved better performance while preserving source data privacy and minimizing communication costs, without requiring the same model structure among different clients.