Learning a Discriminative Null Space for Person Re-identification CVPR 2016

运用KNFST进行降维，用于行人再识别。

原理部分  NFST KNFST

实验部分

数据库

VIPeR: 论文提供了提取特征后的数据。

VIPeR contains 632 identities and each has two images captured outdoor from two views with distinct view angles. All images are scaled to 128 × 48 pixels。

632 people’s images are randomly divided into two equal halves, one for training and the other for testing. This is repeated for 10 times and the averaged performance is reported.  A和B两个场景中各选择一半作为训练集合，另外一半作为测试集合。

A，B的训练集合集成在一起形成训练集。A的测试作为gallary, B的测试作为probe.  最终通过训练得到对特征空间的投影矩阵（隐式的，核方法），计算样本在低维空间中的投影特征，然后利用欧式距离进行识别，评估采用累积匹配曲线， CMC Cumulated Matching Characteristics。

By default the recently proposed Local Maximal Occurrence (LOMO) features [22] are used for person representation. The descriptor has 26,960 imensions. To test our method’s ability to fuse different representations, we also consider another histogram based image descriptor proposed in [24]. These include
colour histogram, HOG and LBP which are concatenated resulting in 5138 dimensions.