Synergistic Cross-Modal Prompt Learning and Structural Knowledge Distillation for Cloth-Changing Person Re-Identification

This repository contains the official PyTorch implementation of our paper:

"Synergistic Cross-Modal Prompt Learning and Structural Knowledge Distillation for Cloth-Changing Person Re-Identification"

• Paper Link: https://doi.org/10.1016/j.knosys.2026.116132

The source code is now publicly available.

📢 News & Updates

• [2026/05/10] The source code has been released.
• The datasets and trained weights are provided via Google Drive.

🛠️ Installation

Please install conda before installing the required packages.

conda create -n malc python=3.8
conda activate malc
pip install -r requirements.txt

📁 Datasets

In the paper, we report results on the following three public cloth-changing person Re-ID datasets:

• PRCC
• LTCC
• VC-Clothes

In addition, we also conducted experiments on:

• Celeb-reID
• Celeb-reID-light

The datasets can be downloaded from Google Drive:

• Dataset Download Link

Please download the datasets and organize them according to the dataset preparation instructions.

Note: When using these datasets, please make sure to properly cite the corresponding original dataset papers and follow their license terms and usage requirements.

🧩 Trained Weights

The trained weights can be downloaded from Google Drive:

• Trained Weights Download Link

🚀 Training

We use the following command for training on LTCC, PRCC, and VC-Clothes:

CUDA_VISIBLE_DEVICES=0 python train_clipreid.py --config_file configs/person/vit_clipreid.yml MODEL.SIE_CAMERA True MODEL.SIE_COE 1.0 MODEL.STRIDE_SIZE '[12, 12]'

For different datasets, please modify the dataset type in the following config file:

configs/person/vit_clipreid.yml

Specifically, when training on PRCC, LTCC, or VC-Clothes, please set the corresponding dataset type in configs/person/vit_clipreid.yml.

Loss Weight Adjustment

If you need to adjust the loss weights, please modify the corresponding settings in:

loss/make_loss.py

Specifically, the loss weights can be adjusted around line 67.

🔍 Evaluation

We use the following command for evaluation:

CUDA_VISIBLE_DEVICES=0 python test_clipreid.py --config_file configs/person/vit_clipreid.yml TEST.WEIGHT 'your_trained_checkpoints_path/ViT-B-16_60.pth'

For PRCC and LTCC, this command can directly report the results under both evaluation settings:

• General / SC
• CC

For VC-Clothes, the same evaluation command is used. However, the dataset type should be changed according to the evaluation setting:

Evaluation Setting	Dataset Type
SC	vcclothes_sc
CC	vcclothes_cc

Please modify the dataset type in configs/person/vit_clipreid.yml before evaluating each setting.

📊 Results

The results on PRCC, LTCC, and VC-Clothes are reported in our paper.

We also conducted additional experiments on Celeb-reID and Celeb-reID-light.

📬 Contact

If you have any questions about this repository or our paper, please feel free to open an issue or contact the authors.

🙏 Acknowledgement

This project is implemented based on PyTorch.

Our code is partially based on the implementations of CLIP-ReID and CAL. We sincerely thank the authors for their excellent work.

• CLIP-ReID: GitHub Repository
• CAL: GitHub Repository

We also thank the authors of related cloth-changing person re-identification works, datasets, and open-source repositories for their valuable contributions to the community.

📄 Citation

If you find this work useful for your research, please consider citing our paper.

@article{ding2026synergistic,
  title={Synergistic cross-modal prompt learning and structural knowledge distillation for cloth-changing person re-identification},
  author={Ding, Yongkang and Wang, Yuxiang and Su, Yiyun and Tian, Yu and Ye, Zi and Jian, Xiangzhou},
  journal={Knowledge-Based Systems},
  volume={345},
  pages={116132},
  year={2026},
  publisher={Elsevier},
  doi={10.1016/j.knosys.2026.116132}
}