In context
By early 2025, unsupervised person re-identification (Re-ID) had become critical for multi-camera systems in robotics and surveillance, where labeling data across camera views is impractical. Existing cluster-based methods often used a single proxy per cluster, struggling with camera-induced variations. This research, published in the Chinese journal Robot, addressed that gap with a camera-aware diverse proxies approach.
What was reported
Researchers from Shenyang University of Chemical Technology and the State Key Laboratory of Robotics (Shenyang Institute of Automation, CAS) proposed a method that constructs multiple proxies per camera class to handle both inter-camera discrepancies and intra-camera differences. The camera multi-proxies include a mean proxy and a hard proxy, which complement each other to tighten intra-camera sample distribution. A camera discrepancy proxy, updated using mean proxies from different camera classes within the same cluster, reduces cross-camera feature gaps.
During training, the method alternates clustering (DBSCAN) and contrastive learning, with specialized losses for multi-proxies and discrepancy proxies. The approach was evaluated on three public datasets, achieving mAP/Rank-1 of 90.4%/96.4% on Market1501, 75.8%/86.5% on DukeMTMC-reID, and 58.2%/82.5% on MSMT17, outperforming prior unsupervised methods.
Why it mattered
This work demonstrated that explicitly modeling camera variations with diverse proxies can significantly improve unsupervised Re-ID, making it more practical for real-world deployments where labeled data are scarce. For industrial automation and multi-robot systems, robust Re-ID enables reliable person tracking across distributed cameras, enhancing security and human-robot interaction without extensive annotation efforts.
“The camera discrepancy proxy and camera multi proxies collaborate with each other to improve intra-class tightness and model performance.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-03-15 · “基于相机感知多样代理的无监督行人重识别”