Deep metric learning (DML) aims to learn a consistent distance embedding where an anchor is closer within the same category than others. It underpins a variety of essential and significant tasks in the development of smart city including face recognition, landmark retrieval, pedestrian detection, person/vehicle re-identification, and so on. Traditional pair-based DML methods try to make full use of the data-to-data relations within a (mini-)batch, but they cannot grasp the data distribution information due to the batch size limitation. On the other hand, proxy-based DML schemes use different proxies to approximate the data distribution. However, the proxies are too sample to represent the intra-category variance. In this paper, we propose a simple but effective method, named soft-instance-label proxy, for embedding learning. It can capture the globe data distribution information while depicting the detailed intra-class data structure. The state-of-the-art empirical results on three public image retrieval benchmarks and two backbone networks demonstrate the superiority of our proposed method. Our Soft-instance-label proxy method can have a Recall$@1$ improvement of 2.4% with Googlenet, largely surpassing the current state-of-art-methods while demonstrating great potential in the development of smart city.

深度度量学习（DML) 旨在学习一致的距离嵌入，其中锚在同一类别中比其他锚更近。它支撑着智慧城市发展中的各种基本和重要任务，包括人脸识别、地标检索、行人检测、人/车重新识别等。传统的基于对的 DML 方法试图充分利用（小）批次内的数据到数据关系，但由于批次大小的限制，它们无法掌握数据分布信息。另一方面，基于代理的 DML 方案使用不同的代理来近似数据分布。然而，代理样本太多，无法代表类别内的方差。在本文中，我们提出了一种简单但有效的方法，称为软实例标签代理，用于嵌入学习。它可以在描绘详细的类内数据结构的同时捕获全球数据分布信息。在三个公共图像检索基准和两个骨干网络上的最新经验结果证明了我们提出的方法的优越性。我们的软实例标签代理方法可以有一个召回$@1$ 使用 Googlenet 提高了 2.4%，大大超过了当前最先进的方法，同时展示了智慧城市发展的巨大潜力。