Multi-level fine-grained center calibration network for unsupervised person re-identification

Person re-identification (ReID) aims to match individuals across different camera views. Unlike traditional supervised methods, unsupervised ReID bypasses the need for costly manual annotations, making it highly desirable for real-world applications. In recent years, clustering-based pseudo-labeling has become a widely used approach in unsupervised person re-identification, achieving state-of-the-art performance on several benchmarks. However, two key limitations remain: (1) Biased Cluster Centers: Hard samples introduce bias into the cluster centers, diminishing the effectiveness of cluster center based contrastive learning. (2) Limitations of Local Features: Existing methods primarily rely on horizontal stripe pooling to extract local features, constraining their capacity to represent sample diversity. To address these limitations, we propose a novel Multi-Level Fine-Grained Center Calibration Network (MFCC) integrating a Fine-Grained Enhanced Feature Extractor and a Center-Guided Feature Calibration module. The Fine-Grained Enhanced Feature Extractor employs a multi-level attention strategy, incorporating low to high level clues, to dynamically identify discriminative regions and extract fine-grained local features. The Center-Guided Feature Calibration module uses a Gaussian Mixture Model (GMM) to identify and calibrate hard samples toward the center of easy samples, resulting in more compact clusters and refined cluster centers. Extensive experiments on two benchmark datasets, Market-1501 and MSMT17, demonstrate the effectiveness of our proposed MFCC framework.