A federated vehicle re-identification benchmark and performance optimization using dual-phase contrastive dynamic aggregation.

Vehicle re-identification is pivotal in advancing intelligent transportation systems and smart city initiatives, but it faces privacy risks due to the central processing of extensive data. To address this, we introduce a federated vehicle re-identification (FV-REID) benchmark with a multi-domain dataset from five sources, evaluation protocols, and a baseline federated-averaging vehicle re-identification method (FVVR). Our studies show that FVVR underperforms compared to traditional models, especially under varied data distribution, and exhibits greater variability. To improve this, we propose a dual-phase contrastive dynamic aggregation (DCDA) method, which adjusts aggregation weights based on model changes during training, allocating greater weights during significant early changes and smaller weights later. This approach addresses data imbalances, allowing clients with smaller datasets to effectively influence the model. Our results demonstrate enhanced performance and stability of the federated vehicle re-identification, contributing significantly to privacy-protected vehicle re-identification.