Abstract

Extending person re-identification (ReID) to a federated scenario has recently drawn attention due to privacy concerns of individuals, but existing methods mostly assume sufficient diversity in pose variations even within a decentralized client. We focus on a more realistic federated-by-camera scenario, where each client corresponds to a single camera and thus captures only a sparse set of poses. To enrich variety, we propose Pose-guided Enriched Feature Learning (PEFL) that explicitly augments pose-diverse samples in the federated ReID scenario. Specifically, a Pose-Extraction Module (PEM) disentangles pose-relevant and pose-irrelevant feature components, where Pose-Relationship Knowledge Distillation (PKD) method helps identify the correct pose and Semantic Consistency Maintenance (SCM) method preserves semantics even with pose changes. In addition, a Compatibility Regularization method ensures the PEM to be compatible with the feature space of the global model. By recombining pose-relevant and -irrelevant components across identities via PEM, our PEFL synthesizes pose-swapped features, thereby largely facilitating contrastive learning of ReID models. Extensive experiments on Market1501 and MSMT17 under the federated-by-camera setting demonstrate that PEFL consistently outperforms federated ReID baselines and their conjunctions with the existing feature augmentation methods; thus achieving state-of-the-art federated ReID performance.

Better AI often begins with better data. For systems designed to identify the same person across different camera views—whether for missing-person searches or public safety—that data is becoming harder to obtain as privacy concerns restrict the sharing of sensitive video. A research team, led by Professor Jae-Young Sim of the Graduate School of Artificial Intelligence at UNIST has developed a federated learning method that overcomes this challenge without transferring raw camera footage to a central server.

Determining whether people captured by different cameras are the same individual—a task known as Person Re-Identification (ReID)—is essential for a wide range of AI vision systems. While federated learning enables models to be trained without sharing sensitive images, distributing training across individual cameras creates an unexpected challenge: each camera observes people from only a limited range of viewpoints.

A camera mounted at one location may capture mostly front-facing views, while another records only side or rear profiles. As a result, local AI models learn from incomplete pose information, making it more difficult for the combined federated model to accurately recognize the same person under different viewing conditions.

To overcome this limitation, the research team developed Pose-guided Enriched Feature Learning (PEFL). Rather than generating or exchanging additional images, PEFL separates visual features related to a person's identity from those associated with pose, then recombines them to synthesize feature representations from viewpoints that were never observed by an individual camera. The approach expands pose diversity during training while preserving each person's distinctive characteristics.

The researchers evaluated PEFL using the widely adopted Market-1501 and MSMT17 benchmarks under a federated-by-camera setting. The method achieved a mean Average Precision (mAP) of 45.9%, improving on the previous state of the art by 4.5 percentage points, while reaching a Rank-1 accuracy of 66.4%.

“We found that the limited range of viewpoints captured by individual cameras significantly reduces the effectiveness of federated person re-identification,” said JooHyung Oh, co-first author of the study. “By enriching pose diversity without sharing raw images, our method enables models to learn more robust identity representations.”

Professor Sim added, "Privacy and performance are often viewed as competing goals in AI. Our work shows they do not have to be. We hope this approach will contribute to distributed vision systems that require both strong recognition performance and robust privacy protection."

The study was co-first authored by JooHyung Oh and Minyoung Oh of UNIST. The research was accepted for presentation at the  IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 , one of the world's leading conferences in computer vision.

Journal Reference

JooHyung Oh, Minyoung Oh, Sung Whan Yoon, and Jae-Young Sim, "Pose-guided Enriched Feature Learning for Federated-by-camera Person Re-identification,"  CVPR '26,  (2026).