Spatial-Temporal Federated Learning for Lifelong Person Re-identification on Distributed Edges

TL;DR

This work addresses data drift in real-world person ReID by introducing FedSTIL, a federated lifelong learning framework that leverages spatial-temporal correlations across distributed edge clients. It combines on-edge lifelong learning with a server-based spatial-temporal knowledge integration, using adaptive layers $ heta_c = B_c \odot \alpha_c + A_c$, task prototypes, and a similarity-weighted aggregation mechanism based on $\mathcal S_{ij}^{(t,t')} = \Pi(\bar{\mathcal P}_i^{(t)}, \bar{\mathcal P}_j^{(t')})$ and forgetting factor $\lambda_f$. The approach yields about 4% improvements in Rank-1 accuracy and 62% reduction in communication cost on a five-dataset, multi-edge ReID benchmark, validating effective cross-edge transfer while preserving privacy. The results demonstrate FedSTIL’s potential for robust, scalable deployment of lifelong ReID on distributed edge devices with privacy constraints and limited communication budgets.

Abstract

Data drift is a thorny challenge when deploying person re-identification (ReID) models into real-world devices, where the data distribution is significantly different from that of the training environment and keeps changing. To tackle this issue, we propose a federated spatial-temporal incremental learning approach, named FedSTIL, which leverages both lifelong learning and federated learning to continuously optimize models deployed on many distributed edge clients. Unlike previous efforts, FedSTIL aims to mine spatial-temporal correlations among the knowledge learnt from different edge clients. Specifically, the edge clients first periodically extract general representations of drifted data to optimize their local models. Then, the learnt knowledge from edge clients will be aggregated by centralized parameter server, where the knowledge will be selectively and attentively distilled from spatial- and temporal-dimension with carefully designed mechanisms. Finally, the distilled informative spatial-temporal knowledge will be sent back to correlated edge clients to further improve the recognition accuracy of each edge client with a lifelong learning method. Extensive experiments on a mixture of five real-world datasets demonstrate that our method outperforms others by nearly 4% in Rank-1 accuracy, while reducing communication cost by 62%. All implementation codes are publicly available on https://github.com/MSNLAB/Federated-Lifelong-Person-ReID

Figures (14)

• Figure 1: The spatial-temporal correlations among the data of different edge clients. The image data captured by different cameras have spatial-temporal correlations, and the edge clients can utilize the spatial-temporal knowledge from others for federated learning to continuously improve their performances.
• Figure 2: The architecture of FedSTIL for federated lifelong person ReID. The distributed edge clients continuously learn from both their local drift data and the relevant spatial-temporal knowledge from other edge clients organized by the parameter server to improve recognition accuracy.
• Figure 3: The architecture of the adaptive layers. The on-edge models continuously learn knowledge from the forthcoming tasks and meanwhile maintain the knowledge from prior tasks. The adaptive layers adaptively balance the tradeoff between local knowledge and global spatial-temporal knowledge.
• Figure 4: The framework of the spatial-temporal knowledge integration on the parameter server. The parameter server can automatically measure the spatial-temporal correlations for edge clients based on their task features. The task-specific knowledge is organized for edge clients for continuously learning.
• Figure 5: The data flow for sampling prototypes into local storage. We dynamically store some identities' prototypes for future rehearsal which are near the corresponding mean center.