Person re-identification across various camera views presents a complex challenge in video surveillance, especially in unsupervised settings where labeled data is unavailable. In such unsupervised person re-identification (re-ID) systems, estimating labels is essential for identifying individuals across different cameras. By using these estimated labels, supervised learning methods can potentially be applied to extract distinctive features. However, this task is complicated by substantial cross-camera variations, which often lead to noisy and inaccurate label predictions. Accurate label estimation, nonetheless, is vital for developing dependable re-ID models. This paper introduces a graph matching technique to address label estimation, where the graph structure is refined using metrics derived from the initially estimated labels to improve accuracy. Additionally, a positive re-weighting strategy has been devised to further refine these intermediate labels, enhancing resilience to errors from mismatches and noisy initial training data. Overall, this paper presents a robust framework for unsupervised person re-identification in real-world scenarios. Through iterative refinement of the graph structure with improved similarity measurements, alongside positive re-weighting and metric learning, the DGM approach enhances label estimation accuracy and mitigates issues from noisy and inaccurate label predictions.