Abstract Part-based methods for person re-identification have been widely studied. In existing part-based methods, although multiple parts are explored, only coarse-grained features of these parts are utilized. Thus, too much fine-grained information is discarded, which limits their ability to extract detailed discriminative features. To tackle this problem, we propose a novel person re-identification network to learn discriminative features across multiple granularities from body regions which are also multi-grained. Specifically, we detect multi-granularity body regions at different stages of a backbone network, and multi-granularity features are learned from body regions with corresponding granularities. To overcome the severe mismatching problem of fine-grained regions and to learn discriminative features, the detection of multi-granularity body regions and the learning of multi-granularity features are jointly optimized. This joint optimization pushes the learned features concentrating on body regions. Moreover, with the body regions well located, the multi-granularity features can be well aligned. Extensive experiments on four popular datasets show that our method is the state-of-the-art in recent years.

中文翻译：

---

从多粒度区域学习多粒度特征以进行人员重新识别

摘要 基于部分的行人重识别方法已被广泛研究。在现有的基于部件的方法中，虽然探索了多个部件，但仅利用了这些部件的粗粒度特征。因此，丢弃了过多的细粒度信息，这限制了它们提取详细判别特征的能力。为了解决这个问题，我们提出了一个新的人重新识别网络，从身体区域的多个粒度中学习区分特征，这些区域也是多粒度的。具体来说，我们在骨干网络的不同阶段检测多粒度的身体区域，并从具有相应粒度的身体区域中学习多粒度特征。为了克服细粒度区域的严重错配问题并学习判别特征，多粒度身体区域的检测和多粒度特征的学习联合优化。这种联合优化将学习到的特征集中在身体区域上。此外，由于身体区域的位置很好，多粒度特征可以很好地对齐。对四个流行数据集的大量实验表明，我们的方法是近年来最先进的。