An Enhanced Metric Learning for Person Reidentification

摘要

Metric learning is a key step in the process of person re-identification. Most metric learning methods are dedicated to proposing new sophisticated algorithms to improve matching accuracy currently. However, re-ranking methods have been ignored, which can be used to re-rank the result of metric learning and improve the matching accuracy of person re-identification effectively. In this paper, we propose a kind of re-ranking method which is based on max-min linear space transformation to deal with the outliers in final distance matrix of traditional metric learning and then we can re-rank the new distance to improve rei-dentification results. Generally, the distances is fixed between each image in gallery with every image in probe when the metric learning method is selected. But the range of the distances value is significantly different from image to image in gallery, we apply max-min criterion to linearly map all values to the specified subspace, which can greatly reduce the error between classes. The effectiveness of our re-ranking method to enhance XQDA algorithm is experimented on two challenging person reidentification databases, CUHK01 and VIPeR. The results show that the accuracy of state-of-the-art is improved.