Taking videos as nodes in a graph, graph matching is an effective technique for unsupervised video-based person re-identification (re-ID). However, most of existing methods are sensitive to noisy training data and mainly only focus on visual content relations between query and gallery videos, which may introduce large amount of false positives. To enhance the robustness to training data and alleviate the visual ambiguity, a Dynamic Hybrid Graph Matching (DHGM) method is proposed, which jointly considers both content and context information for person re-ID in an iterative manner. The content relations between video nodes are obtained by metric learning, based on which the context relation is acquired by encoding the bidirectional feature of each probe node relative to its graph neighbors. The model is iteratively updated during the process of graph construction for promoted distance measurement and further better matching performance. Experimental results on the PRID 2011 and iLIDS-VID datasets demonstrate the superiority of the DHGM.

中文翻译：

---

基于无监督视频的人重新识别的动态混合图匹配

将视频作为图中的节点，图匹配是一种基于视频的无监督行人重新识别（re-ID）的有效技术。然而，现有的大多数方法对嘈杂的训练数据很敏感，并且主要只关注查询和图库视频之间的视觉内容关系，这可能会引入大量的误报。为了增强对训练数据的鲁棒性并减轻视觉模糊性，提出了一种动态混合图匹配（DHGM）方法，该方法以迭代的方式联合考虑内容和上下文信息以进行人员重识别。视频节点之间的内容关系是通过度量学习获得的，在此基础上，通过编码每个探针节点相对于其图邻居的双向特征来获取上下文关系。该模型在图构建过程中迭代更新，以促进距离测量和更好的匹配性能。PRID 2011 和 iLIDS-VID 数据集的实验结果证明了 DHGM 的优越性。