Awesome

MAR: MultilAbel Reference Learning

<img src="img/intro.png" width="400"/> <img src="img/fig3.png" width="450"/>

This repo contains the source code for our CVPR'19 work Unsupervised person re-identification by soft multilabel learning (the paper and the supplementary material is available). Our implementation is based on Pytorch. In the following is an instruction to use the code to train and evaluate the MAR model on the Market-1501 dataset.

Prerequisites

1. Pytorch 1.0.0
2. Python 3.6+
3. Python packages: numpy, scipy, pyyaml/yaml, h5py
4. [Optional] MATLAB, if you need to customize used datasets.

Data preparation

If you simply want to run the demo code without further modification, you might skip this step by downloading all required data from BaiduPan with password "tih8", and put all of them into /data/. Alternatively, you can find processed MSMT17 here.

1. Pretrained model

   Please find the pretrained model (pretrained using softmax loss on MSMT17) in BaiduPan (password: tih8) or GoogleDrive. After downloading pretrained_MSMT17.pth, please put it into /data/.

2. Target dataset

   Download the Market-1501 dataset, and unzip it into /data. After this step, you should have a folder structure:

   • data
     • Market-1501-v15.09.15
       • bounding_box_test
       • bounding_box_train
       • query

   Then run /data/construct_dataset_Market.m in MATLAB. If you prefer to use another dataset, just modify the MATLAB code accordingly. The processed Market-1501 and DukeMTMC-reID are available in BaiduPan.

3. Auxiliary (source) dataset

   Download the MSMT17 dataset, and unzip it into /data. After this step, you should have a folder structure:

   • data
     • MSMT17_V1
       • train
       • test
       • list_train.txt
       • list_query.txt
       • list_gallery.txt

   Then run /data/construct_dataset_MSMT17.m in MATLAB. If you prefer to use another dataset, just modify the MATLAB code accordingly. Again, the processed MSMT17 is available in BaiduPan and Mega.

Run the code

Please enter the main folder, and run

python src/main.py --gpu 0,1,2,3 --save_path runs/debug

where "0,1,2,3" specifies your gpu IDs. If you are using gpus with 12G memory, you need 4 gpus to run in the default setting (batchsize=368). If you set a small batch size, please do not forget to lower the learning rate as the gradient would be stronger for a smaller batch size. Please also note that since I load the whole datasets into cpu memory to cut down IO overhead, you need at least 40G cpu memory. Hence I recommend you run it on a server.

Main results

Reference

If you find our work helpful in your research, please kindly cite our paper:

Hong-Xing Yu, Wei-Shi Zheng, Ancong Wu, Xiaowei Guo, Shaogang Gong and Jian-Huang Lai, "Unsupervised person re-identification by soft multilabel learning", In CVPR, 2019.

bib:

@inproceedings{yu2019unsupervised,
  title={Unsupervised Person Re-identification by Soft Multilabel Learning},
  author={Yu, Hong-Xing and Zheng, Wei-Shi and Wu, Ancong and Guo, Xiaowei and Gong, Shaogang and Lai, Jianhuang},
  year={2019},
  booktitle={IEEE International Conference on Computer Vision and Pattern Recognition (CVPR)},
}

Contact

If you have any problem please email me at xKoven@gmail.com I may not look at issues.