FAIM: Clothes-Changing Person Re-Identification with Feasibility-Aware Intermediary Matching

FAIM is a clothes-changing person re-id method that leverages intermediaries to improve the identity matching results in clothes-changing scenarios. For more details, please refer to the paper.

Introduction

We propose an intermediary matching method, using extra intermediaries to improve the clothes-changing matching results when clothes-irrelevant features are of low quality.

Dataset Preparation

We mainly use three clothes-changing re-id benchmarks for training and evaluation: LTCC, PRCC and DeepChange.

Environment Setup

• Python 3.8, PyTorch 1.12
• Install the dependencies:

pip install -r requirements.txt

• To run GNN re-ranking method, you need to manually install the gnn reranking tools

Training

Take LTCC dataset as an example

• Setup the configuration files at configs/res50_cels_cal_reliability_ltcc.yaml. Replace DATA.ROOT to your own root directory containing LTCC dataset. Replace OUTPUT to your own output directory.
• Run the corresponding lines in scripts.sh

Testing

Take the testing of LTCC dataset as an example

• Setup the test configurations at configs/res50_cels_cal_reliability_test.yaml.
  1. Replace DATA.ROOT to your own root directory containing LTCC dataset. Replace OUTPUT to your own output directory.
  2. For FAIM implemented with K-reciprocal neighboring, set TEST.RERANKING=1. For FAIM implemented with GNN, set TEST.RERANKING=2.
• Run the corresponding lines in test_scripts.sh. Change the --resume to your own checkpoint path.

Some Visualizations

Citations

If you find our paper and codebase helpful, please consider citing

@misc{zhao2024clotheschanging,
      title={Clothes-Changing Person Re-Identification with Feasibility-Aware Intermediary Matching}, 
      author={Jiahe Zhao and Ruibing Hou and Hong Chang and Xinqian Gu and Bingpeng Ma and Shiguang Shan and Xilin Chen},
      year={2024},
      eprint={2404.09507},
      archivePrefix={arXiv},
      primaryClass={id='cs.CV' full_name='Computer Vision and Pattern Recognition' is_active=True alt_name=None in_archive='cs' is_general=False description='Covers image processing, computer vision, pattern recognition, and scene understanding. Roughly includes material in ACM Subject Classes I.2.10, I.4, and I.5.'}
}

Acknowledgements

This code implementation is based on Simple-CCReID.