This repository provides the evaluation codes for the MARS dataset
Matlab C C++
Switch branches/tags
Nothing to show
Clone or download
Latest commit 3a91bbc Jun 13, 2017
Failed to load latest commit information.
CM_Curve Add files via upload Aug 9, 2016
KISSME Add files via upload Aug 9, 2016
LOMO_XQDA Add files via upload Aug 9, 2016
info Add files via upload Aug 9, 2016
utils fix the problem of pca Jun 13, 2017 Update Aug 9, 2016
test_mars.m Add files via upload Aug 9, 2016


This code provides evaluation procedure of the MARS dataset. Please kindly cite the Arxiv paper if you use this dataset.

Liang Zheng*, Zhi Bie*, Yifan Sun*, Jingdong Wang, Chi Su, Shengjin Wang, Qi Tian, "MARS: A Video Benchmark for Large-Scale Person Re-identification", ECCV, 2016. (* equal contribution)

This code uses the 1024-dim IDE descriptor [1] and KISSME [2] and XQDA [3] distance metrics. To run this code, one should follow the three steps below.

  1. Download the pre-computed IDE feature: or Unzip it in the root folder.

  2. Run "test_mars.m".

If you want to try your own descriptor or to learn new features, you should do as follows.

  1. Download the dataset: or Training should be done with images in folder "bbox_train".

  2. Bounding box feature extraction should follow the order specified in "root/info/test_name.txt" and "root/info/train_name.txt." The newly extracted feature should be loaded in line 19-20 in "root/test_mars.m"

If you have any suggestions or comments, please email me at


[1] L. Zheng et al. Person Re-identification in the Wild. Arxiv, 2016.

[2] S. Liao et al. Person re-identification by local maximal occurrence representation and metric learning. CVPR 2015.

[3] M. Kostinger et al. Large scale metric learning from equivalence constraints. CVPR 2012.