Research on the application of vehicle re-identification to video surveillance has attracted increasingly growing attention. Existing methods are associated with the difficulties of distinguishing different instances of the same car model owing to the incapability of recognizing subtle differences among these instances and the possibility that a subtle difference may lead to incorrect results of ranking. In this paper, a discriminative fine-grained network for vehicle re-identification based on a two-stage re-ranking framework is proposed to address these issues. This discriminative fine-grained network (DFN) is composed of fine-grained and Siamese networks. The proposed hybrid network can extract discriminative features of the vehicle instances with subtle differences. The Siamese network is rather suitable for general object re-identification using two streams of the network, while the fine-grained network is capable of detecting subtle differences. The proposed two-stage re-ranking method allows obtaining a more reliable ranking list by using the Jaccard metric and merging the first and second re-ranking lists, where the latter list is formed using the sample mean feature. Experimental results on the VeRi-776 and VehicleID datasets show that the proposed method achieves the superior performance compared to the state-of-the-art methods used in vehicle re-identification.

关于车辆重新识别在视频监控中的应用的研究已引起越来越多的关注。由于无法识别这些实例之间的细微差异以及细微差异可能导致错误的排名结果的可能性，现有方法与区分同一汽车模型的不同实例的困难相关。为了解决这些问题，本文提出了一种基于两阶段重新排序框架的用于车辆重新识别的判别性细粒度网络。此区分性细粒度网络（DFN）由细粒度和暹罗网络组成。提出的混合网络可以提取具有细微差异的车辆实例的判别特征。连体网络非常适合使用网络的两个流来进行一般对象的重新标识，而细粒度网络则能够检测到细微的差异。所提出的两阶段重新排名方法允许通过使用Jaccard度量并将第一和第二重新排名列表合并来获得更可靠的排名列表，其中后一个列表是使用样本均值特征形成的。在VeRi-776和VehicleID数据集上的实验结果表明，与用于车辆重新识别的最新方法相比，该方法具有更好的性能。所提出的两阶段重新排名方法允许通过使用Jaccard度量并将第一和第二重新排名列表合并来获得更可靠的排名列表，其中后一个列表是使用样本均值特征形成的。在VeRi-776和VehicleID数据集上的实验结果表明，与用于车辆重新识别的最新方法相比，该方法具有更好的性能。所提出的两阶段重新排名方法允许通过使用Jaccard度量并将第一和第二重新排名列表合并来获得更可靠的排名列表，其中后一个列表是使用样本均值特征形成的。在VeRi-776和VehicleID数据集上的实验结果表明，与用于车辆重新识别的最新方法相比，该方法具有更好的性能。