This paper proposes a Multiple Granularity Network (MGN) for person re-identification (Re-ID), which integrates global and local feature learning with different granularities. The MGN is a multi-branch deep network consisting of one global branch and two local branches. The global branch learns global feature representations, while the local branches learn local feature representations with varying granularities by partitioning images into stripes. The method achieves state-of-the-art performance on mainstream Re-ID datasets such as Market-1501, DukeMTMC-reID, and CUHK03. On Market-1501, the method achieves Rank-1/mAP of 96.6%/94.2% in single query mode. The MGN uses a combination of softmax loss and triplet loss for training, and is fully end-to-end. The method outperforms existing approaches by a large margin, demonstrating the effectiveness of combining global and local features with different granularities for Re-ID. The MGN architecture is designed to learn more discriminative features by leveraging the hierarchical structure of the ResNet-50 backbone. The method is evaluated on multiple datasets and shows robust performance across different scenarios. The results indicate that the MGN is a powerful approach for person Re-ID, achieving high accuracy and robustness in challenging conditions.This paper proposes a Multiple Granularity Network (MGN) for person re-identification (Re-ID), which integrates global and local feature learning with different granularities. The MGN is a multi-branch deep network consisting of one global branch and two local branches. The global branch learns global feature representations, while the local branches learn local feature representations with varying granularities by partitioning images into stripes. The method achieves state-of-the-art performance on mainstream Re-ID datasets such as Market-1501, DukeMTMC-reID, and CUHK03. On Market-1501, the method achieves Rank-1/mAP of 96.6%/94.2% in single query mode. The MGN uses a combination of softmax loss and triplet loss for training, and is fully end-to-end. The method outperforms existing approaches by a large margin, demonstrating the effectiveness of combining global and local features with different granularities for Re-ID. The MGN architecture is designed to learn more discriminative features by leveraging the hierarchical structure of the ResNet-50 backbone. The method is evaluated on multiple datasets and shows robust performance across different scenarios. The results indicate that the MGN is a powerful approach for person Re-ID, achieving high accuracy and robustness in challenging conditions.