Spatial-Temporal Person Re-identification

Spatial-Temporal Person Re-identification

Most of current person re-identification (ReID) methods neglect aspatial-temporal constraint. Given a query image, conventional methods computethe feature distances between the query image and all the gallery images andreturn a similarity ranked table.When the gallery database is very large inpractice, these approaches fail to obtain a good performance due to appearanceambiguity across different camera views. In this paper, we propose a noveltwo-stream spatial-temporal person ReID (st-ReID) framework that mines bothvisual semantic information and spatial-temporal information. To this end, ajoint similarity metric with Logistic Smoothing (LS) is introduced to integratetwo kinds of heterogeneous information into a unified framework. To approximatea complex spatial-temporal probability distribution, we develop a fastHistogram-Parzen (HP) method. With the help of the spatial-temporal constraint,the st-ReID model eliminates lots of irrelevant images and thus narrows thegallery database. Without bells and whistles, our st-ReID method achievesrank-1 accuracy of 98.1\% on Market-1501 and 94.4\% on DukeMTMC-reID, improvingfrom the baselines 91.2\% and 83.8\%, respectively, outperforming all previousstate-of-the-art methods by a large margin.

时空人重新识别

当前大多数人重新识别（ReID）方法都忽略了时空约束。给定查询图像，常规方法计算查询图像与所有图库图像之间的特征距离，并返回相似性排序表。.. 实际上，当图库数据库很大时，由于不同摄影机视图之间的模棱两可，这些方法无法获得良好的性能。在本文中，我们提出了一种新颖的两流时空人ReID（st-ReID）框架，该框架同时挖掘视觉语义信息和时空信息。为此，引入了具有Logistic平滑（LS）的联合相似性度量，以将两种异构信息集成到一个统一的框架中。为了近似复杂的时空概率分布，我们开发了一种快速的直方图-Parzen（HP）方法。借助时空约束，st-ReID模型消除了许多不相关的图像，从而缩小了图库数据库的范围。在没有花哨的情况下，我们的st-ReID方法在Market-1501和94上实现了98.1％的1级精度。 （阅读更多）