Generalizing A Person Retrieval Model Hetero- and Homogeneously

================================================================

Code for Generalizing A Person Retrieval Model Hetero- and Homogeneously (ECCV 2018). [paper]

Preparation

Requirements: Python=3.6 and Pytorch=0.4.0

1. Install Pytorch

2. Download dataset

   • reid_dataset [GoogleDriver]

   • The reid_dataset including Market-1501 (with CamStyle), DukeMTMC-reID (with CamStyle), and CUHK03

   • Unzip reid_dataset under 'HHL/data/'

CamStyle Generation

You can train CamStyle model and generate CamStyle imgaes with stargan4reid

Training and test domain adaptation model for person re-ID

1. Baseline

# For Duke to Market-1501
python baseline.py -s duke -t market --logs-dir logs/duke2market-baseline
# For Market-1501 to Duke
python baseline.py -s market -t duke --logs-dir logs/market2duke-baseline

2. HHL

# For Duke to Market-1501
python HHL.py -s duke -t market --logs-dir logs/duke2market-HHL
# For Market-1501 to Duke
python HHL.py -s market -t duke --logs-dir logs/market2duke-HHL

Results

<table> <tr> <td></td> <td colspan="2">Duke to Market</td> <td colspan="2">Market to Duke</td> </tr> <tr> <td>Methods</td> <td>Rank-1</td> <td>mAP</td> <td>Rank-1</td> <td>mAP</td> </tr> <tr> <td>Baseline</td> <td>44.6</td> <td>20.6</td> <td>32.9</td> <td>16.9</td> </tr> <tr> <td>HHL</td> <td>62.2</td> <td>31.4</td> <td>46.9</td> <td>27.2</td> </tr> </table>

References

• [1] Our code is conducted based on open-reid

• [2] StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation , CVPR 2018

• [3] Camera Style Adaptation for Person Re-identification. CVPR 2018.

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{zhong2018generalizing,
title={Generalizing A Person Retrieval Model Hetero- and Homogeneously},
author={Zhong, Zhun and Zheng, Liang and Li, Shaozi and Yang, Yi},
booktitle ={ECCV},
year={2018}
}

Contact me

If you have any questions about this code, please do not hesitate to contact me.

Zhun Zhong