Vehicle re-identification is a crucial research direction in computer vision for constructing intelligent transportation systems and smart cities. However, privacy concerns pose significant challenges, such as personal information leakage and potential risks of data sharing. To address these challenges, we propose a federated vehicle re-identification (FV-REID) benchmark that protects vehicle privacy while exploring re-identification performance. The benchmark includes a multi-domain dataset and a federated evaluation protocol that allows clients to upload model parameters to the server without sharing data. We also design a baseline federated vehicle re-identification method called FVVR, which employs federated-averaging to facilitate model interaction. Our experiments on the FV-REID benchmark reveal that (1) the re-identification performance of the FVVR model is typically weaker than that of non-federated learning models and is prone to significant fluctuations and (2) the difference in re-identification performance between the FVVR model and the non-federated learning model would be more pronounced on a small-scale client dataset compared to a large-scale client dataset.
CITATION STYLE
Huang, L., Zhao, Q., Zhou, L., Zhu, J., & Zeng, H. (2023). FV-REID: A Benchmark for Federated Vehicle Re-identification. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 14463 LNCS, pp. 395–406). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-981-99-8565-4_37
Mendeley helps you to discover research relevant for your work.