Metric learning is one of the important ways to improve the person re-identification (ReID) accurate, of which triplet loss is the most effect metric learning method. However, triplet loss only ranks the extracted feature at the end of the network, in this paper, we propose a multilevel metric rank match (MMRM) method, which ranks the extracted feature on multilevel of the network. At each rank level, the extracted features are ranked to find the hard sample pairs and the back transfer triplet loss. Each rank level has different penalize value to adjust the network, in which the value is bigger with the deeper level of the whole network. Experiment results on CUHK03, Market1501 and DukeMTMC datasets indicate that The MMRM algorithm can outperform the previous state-of-the-arts.

Chao Wang is now a Ph.D. candidate with the Institute of Machine Learning and Systems Biology, School of Electronics and Information Engineering, Tongji University, China. He received the B.S. degree from Anhui University, China, in 2006, and the M.S. degree from University of Science and Technology of China, in 2009. From Nov. 2009 to Dec. 2010, he worked as Research Associate in Nanyang Technological University. From Jan. to May 2011, he worked as Research Associate in Jacobs University, Bremen Germany. From June 2011 to Aug. 2012, he worked as Research Associate in Shenzhen Institute of advanced technology, Chinese Academy of Sciences. From Sept. 2012 to Mar. 2013, he worked as Engineer in Huawei Technology Co., Ltd. He now works at Suzhou University. His research focuses on pattern recognition, deep learning and image processing.