FedSH: Towards Privacy-preserving Text-based Person Re-Identification

The implementation of paper [FedSH: Towards Privacy-preserving Text-based Person Re-Identification]

Requirements

torch >= 1.7.0
yaml
omegaconf
visdom
Pillow 8.2.0

Data preparation

1. CUHK-PEDES

Download the CUHK-PEDES dataset from here

Organize them in data folder as follows:

|-- data/
|   |-- <CUHK-PEDES>/
|       |-- imgs
            |-- cam_a
            |-- cam_b
            |-- CUHK01
            |-- CUHK03
            |-- Market
|       |-- reid_raw.json
|-- fllib/

2. ICFG-PEDES

Download the ICFG-PEDES dataset from here

Organize them in data folder as follows:

|-- data/
|   |-- <ICFG-PEDES>/
|       |-- imgs
            |-- test
            |-- train 
|       |-- ICFG_PEDES.json
|-- fllib/

3. Data preprocessing

then run the process_CUHK_data.py and process_ICFG_data.py in SSAN

How to Run

• if you have opened the visualization, you should start the visdom first.

  call start_visdom.bat
  

  You can see the training lines in localhost:8097(Default)

• Then start to train directly

  python train.py
  

• After training done, you can test your model by run:

  python test.py
  

Citation

If you find FedSH useful in your work, please consider staring 🌟 this repo and citing 📑 our paper:

@ARTICLE{10310121,
  author={Ma, Wentao and Wu, Xinyi and Zhao, Shan and Zhou, Tongqing and Guo, Dan and Gu, Lichuan and Cai, Zhiping and Wang, Meng},
  journal={IEEE Transactions on Multimedia}, 
  title={FedSH: Towards Privacy-Preserving Text-Based Person Re-Identification}, 
  year={2024},
  volume={26},
  number={},
  pages={5065-5077},
  keywords={Semantics;Training;Task analysis;Privacy;Visualization;Federated learning;Servers;Text-based Person ReID;Cross-modal Retrieval;Federated Learning;Multi-granularity Representation},
  doi={10.1109/TMM.2023.3330091}}

Copyright

• The code is provided by Wentao Ma and Xinyi Wu from NUDT. If you have any question, please contact wtma@nudt.edu.cn or wuxinyi17@nudt.edu.cn.

Acknowledgments

Our code is based on SSAN and FLLIB.