Content area
Unsupervised person re-identification (Re-ID) is a critical and challenging task in computer vision. It aims to identify the same person across different camera views or locations without using any labeled data or annotations. Most existing unsupervised Re-ID methods adopt a clustering and fine-tuning strategy, which alternates between generating pseudo-labels through clustering and updating the model parameters through fine-tuning. However, this strategy has two major drawbacks: (1) the pseudo-labels obtained by clustering are often noisy and unreliable, which may degrade the model performance; and (2) the model may overfit to the pseudo-labels and lose its generalization ability during fine-tuning. To address these issues, we propose a novel method that integrates silhouette coefficient-based label correction and contrastive loss regularization based on loose–tight cluster guidance. Specifically, we use silhouette coefficients to measure the quality of pseudo-labels and correct the potential noisy labels, thereby reducing their negative impact on model training. Moreover, we introduce a new contrastive loss regularization term that consists of two components: a cluster-level contrast loss that encourages the model to learn discriminative features, and a regularization loss that prevents the model from overfitting to the pseudo-labels. The weights of these components are dynamically adjusted according to the silhouette coefficients. Furthermore, we adopt Vision Transformer as the backbone network to extract more robust features. We conduct extensive experiments on several public datasets and demonstrate that our method achieves significant improvements over the state-of-the-art unsupervised Re-ID methods.