EMGC : An Efficient Unsupervised Person Re‐Identification Method Based on Multigrained Complementary Features

What the paper says

Unsupervised person re‐identification (Re‐ID) aims to learn discriminative features from unlabelled data using pseudo‐labels for pedestrian retrieval. However, existing methods fail to fully leverage fine‐grained clues from multigranularity features when optimising pseudo‐labels, which limits their performance. To address this issue, this paper proposes an Efficient Multigranularity Complementary feature‐based unsupervised person Re‐ID method (EMGC), which consists of two core modules: a Multigranularity Token Propagation module (MGTP) and a Complementary Feature Label Optimisation module (CFLO). The MGTP is built upon an improved Vision Transformer (ViT) and enhances feature representation capability while reducing computational complexity through a Multigranularity Architecture (MGA) and an Efficient Token Propagation strategy (ETP). Specifically, MGA captures both global and local features, while ETP reduces redundancy and improves computational efficiency via dynamic token selection. The CFLO introduces a Global and Partial Feature Complementary Pseudo‐Label Optimisation framework (GPFC‐PLO), which leverages fine‐grained information from local features to improve pseudo‐label quality, mitigate noise interference and enhance robustness. Extensive experiments on three person Re‐ID datasets demonstrate the effectiveness of EMGC. The results show that it outperforms current state‐of‐the‐art unsupervised methods in both performance and inference speed, significantly narrowing the performance gap with supervised approaches.