@article{Li2025, 
author = {Mingkun Li and Peng Xu},
title = {Silhouette-driven contrastive learning with hierarchical neighborhood structure for unsupervised clothes change person re-identification},
year = {2025},
journal = {CAAI Artificial Intelligence Research},
volume = {4},
pages = {9150050},
keywords = {unsupervised learning, contrastive learning, clothes change person re-ID, silhouette-driven},
url = {https://www.sciopen.com/article/10.26599/AIR.2025.9150050},
doi = {10.26599/AIR.2025.9150050},
abstract = {This paper addresses a highly challenging yet critical task: unsupervised person re-identification (Re-ID) under clothing changes. Unfortunately, existing unsupervised person Re-ID methods are mainly designed for short-term scenarios that rely heavily on RGB cues, thus failing to exploit effective features invariant to clothing changes. To tackle this limitation, we propose a novel and efficient approach, called silhouette-driven contrastive learning with hierarchical neighborhood structure (SiCL-HN). Our framework integrates RGB and silhouette information within a contrastive learning paradigm, where silhouettes provide clothing-invariant shape cues that complement RGB features, and constructs a hierarchical neighborhood structure among instances to learn features robust to clothing variations. We conduct extensive experiments to evaluate our proposed approach against state-of-the-art unsupervised person Re-ID methods on six representative long-term datasets (LTCC, PRCC, VC-Clothes, Celeb-ReID, Celeb-ReID-Light, and DeepChange), and observe consistent improvements on mAP and Rank-1; for example, on DeepChange our method surpasses SpCL by +3.5 mAP and +8.1 Rank-1. Experimental results demonstrate that our approach significantly outperforms existing unsupervised Re-ID methods. Our code is available at https://github.com/MingkunLishigure/SiCL.}
}