基于双向关系相似度函数学习的行人再识别

Learning Bidirectional Relationship Similarity Function for Person Re-Identification

当前的行人再识别在度量学习上采用马氏距离相似度函数,该相似度函数只与特征差分空间有关,忽略了一对行人图像中每个个体的外观特征,针对上述问题,提出了通过学习一个双向关系相似度函数(Bidirectional Relationship Similarity Function,BRSF),来计算一对行人图像的相似度.BRSF不但描述了一对行人图像的互相关关系,而且关联了一对行人图像的自相关关系.该文利用KISSME(Keep It Simple and Straightforward Metric)算法的思想进行相似度函数学习,把一对样本特征的自相关关系和互相关关系用高斯分布来表示,通过把最终高斯分布的比值转换为BRSF的形式,得到一个对背景、视角、姿势的变化具有鲁棒性的相似度函数.在VIPe R,QMUL GRID两个行人再识别数据集上的实验结果表明,本文算法具有较高的识别率,其中在VIPe R数据集上,Rank1达到了53.21%.

These dominant algorithms to learn a similarity is the metric learning that learns a Mahalanobis Similarity Function（MSF） to estimate the similarity of a pair of persons. However, the MSF only projects a pair of persons into feature difference space and ignores the appearance of each individual. In this study, we proposed to learn a Bidirectional Relationship Similarity Function（BRSF） that greatly strengthens the modeling ability of the similarity function. BRSF not only represents the cross correlation relationship of a pair of persons, but also describes the auto correlation relationship.We use the ideal of the Keep It Simple Straightforward Metric（KISSME） algorithm to learn a similarity function.Specifically, the auto correlation relationship and cross correlation relationship of a pair of sample features are expressed by Gaussian distribution. Finally, by converting the ratio of the final Gaussian distribution into the form of BRSF, we get a similarity function which is robust to the change of background, viewpoint, and posture. The proposed method is demonstrated on two public benchmark datasets including VIPe R and QMUL GRID, and experimental results show that the proposed method achieves excellent re-identification rates compared with other similar algorithms. Moreover, the reidentification results on the VIPe R dataset with half of dataset sampled as training samples are quantitatively analyzed,and the performance of the proposed method achieves a 53.21% at Rank1（represents the correct matched pair）.