A Discriminatively Learned CNN Embedding for Person Re-identification

In this package, we provide our training and testing code written in Matconvnet for the paper [A Discriminatively Learned CNN Embedding for Person Re-identification] (https://arxiv.org/abs/1611.05666).

We also include matconvnet-beta23 which has been modified for our paper. All codes have been test on Ubuntu14.04 and Ubuntu16.04 with Matlab R2015b.

This code is ONLY released for academic use.

• Xuanyi Dong also realizes our paper in Caffe. Alternatively, you can run the code with Caffe.
• Zhun Zhong provides a extensive Caffe baseline code. You may check it.

~What's new: We add the data preparation and evaluation codes for CUHK03.

~What's new: We make the code of model structure more easy to follow.

~What's new: We provide a better code for extract feature.

~What's new: We provide a faster evaluation code for Market-1501.

Installation

1. Clone this repo

   git clone https://github.com/layumi/2016_person_re-ID.git
   cd 2016_person_re-ID
   mkdir data

2. Download the pretrained model.

   This model is ONLY released for academic use. You can find the pretrained model in GoogleDriver or [BaiduYun] (https://pan.baidu.com/s/1miG2OpM). Download and put them into the ./data.

   BaiduYun sometime changes the link. If you find the url fail, you can contact me to update it.

3. Compile matconvnet (Note that I have included my Matconvnet in this repo, so you do not need to download it again. I has changed some codes comparing with the original version. For example, one of the difference is in /matlab/+dagnn/@DagNN/initParams.m. If one layer has params, I will not initialize it again, especially for pretrained model.)

   You just need to uncomment and modify some lines in gpu_compile.m and run it in Matlab. Try it~ (The code does not support cudnn 6.0. You may just turn off the Enablecudnn or try cudnn5.1)

   If you fail in compilation, you may refer to http://www.vlfeat.org/matconvnet/install/

Dataset

• Download Market1501 Dataset. The photos are taken in Tsinghua University.

• DukeMTMC-reID is a larger dataset in the same format of Market1501. The photos are taken in Duke University. You can download it from DukeMTMC-reID Dataset. We also upload the result to DukeMTMC-reID leaderboard.

• If you want to rehearsal our result on CUHK03 Dataset, you can simply change the number of kernel from 751 to 1367 in resnet52_market.m and recreate net.mat. Because there are 751 IDs in Market-1501 while 1367 training identities are in CUHK03. More information can be found in cuhk03-prepare-eval dir. We add the data preparation and evaluation codes for CUHK03.

• Training dataset for Oxford5k (http://cmp.felk.cvut.cz/cnnimageretrieval/)

Test

1. Run test/test_gallery_query_crazy.m to extract the features of images in the gallery and query set. They will store in a .mat file. Then you can use it to do evaluation.
2. Evaluate feature on the Market-1501. Run evaluation/zzd_evaluation_res_faster.m. You can get the following Single-query Result.

Methods	Rank@1	mAP
Ours* (SQ)	80.82%	62.30%
Ours* (MQ-avg)	86.67%	70.16%
Ours* (MQ-max)	86.76%	70.68%
Ours* (MQ-max+rerank)	86.67%	72.55%

*Note that the result is slightly higher than the result reported in our paper.

*For multi-query result, you can use evaluation/zzd_evaluation_res_fast.m . It is slower than evaluation/zzd_evaluation_res_faster.m since it need to extract extra features. (The evaluation code is modified from the Market-1501 Baseline Code)

FQA

1. What is multi-query setting?

Actually, we can get a sequence of the query under one camera instead of one image. Then we can use every image in this sequence to extract a query mean feature (mean of feature extracted from several images). We call it multi-query. If we use this feature to do person retrieval, we usually get a better result. But it use additional images (in 'Market-1501/gt_bboxes'). You can find more detail in the original paper.

Train

1. Add your dataset path into prepare_data.m and run it. Make sure the code outputs the right image path.

2. Run train_id_net_res_2stream.m to have fun.

Citation

Please cite this paper in your publications if it helps your research:

@article{zheng2016discriminatively,
  title={A Discriminatively Learned CNN Embedding for Person Re-identification},
  author={Zheng, Zhedong and Zheng, Liang and Yang, Yi},
  doi={10.1145/3159171},
  journal={ACM Transactions on Multimedia Computing Communications and Applications},
  year={2017}
}

Acknowledge

Thanks for Xuanyi Dong to realize our paper in Caffe.

Thanks for Weihang Chen to report the bug in prepare_data.m.

Related Repos

1. Person re-ID with GAN
2. Pedestrian Alignment Network