Video-Level Language-Driven Video-Based Visible-Infrared Person Re-Identification

This is the official repository for VLD, a novel framework designed for Video-Based Visible-Infrared Person Re-Identification.

---

🚀 Getting Started

📁 Dataset

This project supports two benchmark datasets for Video-based Visible-Infrared Person Re-Identification: VCM and BUPT.

• 📂 VCM Dataset Download & Processing Instructions

• 📂 BUPT Dataset Download & Processing Instructions

🏋️‍♂️ Training

To train VLD on the VCM dataset:

CUDA_VISIBLE_DEVICES=0 python main.py --mode train --dataset vcm  --pid_num 500 --output_path logs/vcm

To train VLD on the BUPT dataset:

CUDA_VISIBLE_DEVICES=0 python main.py --mode train --dataset bupt  --pid_num 1074 --output_path logs/bupt

🚗Evaluation

To evaluate the model on the VCM dataset:

CUDA_VISIBLE_DEVICES=0 python main.py --mode test --dataset vcm   --pid_num 500   --resume_test_path logs/vcm/models  --output_path logs/vcm_test

To evaluate the model on the BUPT dataset:

CUDA_VISIBLE_DEVICES=0 python main.py --mode test --dataset bupt  --pid_num 1074  --resume_test_path logs/bupt/models --output_path logs/bupt_test

💾 Pretrained Models & Training Logs

We provide pretrained weights and training logs for both datasets:

🔹 VCM Dataset

• 📦 Pretrained Model (VCM)
• 📑 Training Log (VCM)

🔹 BUPT Dataset

• 📦 Pretrained Model (BUPT)
• 📑 Training Log (BUPT)

Please download the .pth model files and place them into the appropriate logs/[dataset]/models/ directory before running evaluation.

Citation

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

@article{li2025video,
  title={Video-Level Language-Driven Video-Based Visible-Infrared Person Re-Identification},
  author={Li, Shuang and Leng, Jiaxu and Kuang, Changjiang and Tan, Mingpi and Gao, Xinbo},
  journal={IEEE Transactions on Information Forensics and Security},
  year={2025},
  publisher={IEEE}
}