.png){kind=icon}
.png){kind=icon}
Return
Object Re-Identification Based on Federated Incremental Subgradient Proximal Optimization
.png){kind=icon}
OA
AI
DOI：10.1109/OJCS.2024.3489875.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Object Re-identification (Object ReID) is one of the key tasks in the field of computer vision. However, traditional centralized ReID methods face challenges related to privacy protection and data storage. Federated learning, as a distributed machine learning framework, can utilize dispersed data for model training without sharing raw data, thereby reducing communication costs and ensuring data privacy. However, the real statistical heterogeneity in federated object re-identification leads to domain shift issues, resulting in decreased performance and generalization ability of the ReID model. Therefore, to address the privacy constraints and real statistical heterogeneity in object re-identification, this article focuses on studying the object re-identification method based on the Federated Incremental Subgradient Proximal(FedISP) framework. FedISP effectively alleviates weight divergence and low communication efficiency issues through incremental sub-gradient proximal methods and ring topology, ensuring stable model convergence and efficient communication. Considering the complexity of ReID scenarios, this article adopts a ViT-based task model to cope with feature skew across clients. Additionally, it defines camera federated scenarios and dataset federated scenarios for problem modeling and analysis. Furthermore, due to the heterogeneous classifiers that clients may have, the approach intergrates personalized layers. In the experiments, instance datasets of two federated scenarios were constructed for model training. The final test results show that FedISP can effectively address the privacy protection and statistical heterogeneity issues faced by ReID.
Keywords:
Cameras
Training
Optimization
Computational modeling
Mathematical models
Federated learning
Topology
Servers
Data privacy
Data models
non-IID data
incremental methods
object re-identification
Journal
I
IF:
8.2
Papers:
411
Citations:
810