Combining features of different scales to learn a more discriminative model is an essential solution for person re-identification (Re-ID) tasks. Most existing multi-scale methods are based on the fusion of features from different scales, which cannot exploit information throughly at each scale and cause gradient chaos in optimizing. To address this problem, in this paper we propose an end-to-end multi-scale constraint network(MSCN) to capture detailed information from multiple scales which can independently train each scale and integrate the features of each scale for prediction. In order to retain more information at different scales, we uniformly divide the feature maps into several parts, and vary the number of parts in different scales, then concatenate all the parts in each scale as the entire feature for training. We use both classification loss and metric loss to optimize the network from different aspects. Extensive experiments on three datasets demonstrate that our method achieves very competitive performance. Especially on the CUHK03 dataset, our approach achieves the state-of-the-art results outperforming the current best method by 2.4%/2.0% in Rank-1/mAP.

结合不同尺度的特征以学习更具区分性的模型是人员重新识别（Re-ID）任务的基本解决方案。现有的大多数多尺度方法都是基于不同尺度特征的融合，这种方法无法在每个尺度上充分利用信息，从而导致优化过程中出现梯度混乱。为了解决这个问题，在本文中我们提出了一种端到端多尺度约束网络（MSCN）来捕获来自多个尺度的详细信息，这些尺度信息可以独立地训练每个尺度并整合每个尺度的特征以进行预测。为了在不同的比例尺上保留更多信息，我们将特征图均匀地划分为几个部分，并以不同的比例尺改变零件的数量，然后将每个比例尺中的所有零件串联为整个训练特征。我们使用分类损失和度量损失从不同方面优化网络。在三个数据集上的大量实验表明，我们的方法取得了非常好的竞争性能。尤其是在CUHK03数据集上，我们的方法在Rank-1 / mAP中获得的最新结果优于当前最佳方法2.4％/ 2.0％。