Exploring Fine-Grained Representation and Recomposition for Cloth-Changing Person Re-Identification

Official PyTorch implementation of "Exploring Fine-Grained Representation and Recomposition for Cloth-Changing Person Re-Identification". (TIFS 2024)

Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, Yanwei Fu

Fudan University, Northwestern Polytechnical University

Getting Started

Environment

Python == 3.8
PyTorch == 1.12.1
faiss-gpu == 1.7.2

Prepare Data

Please download cloth-changing person re-identification datasets and place them in any path DATASET_ROOT:

DATASET_ROOT
    └─ LTCC-reID or Celeb-reID or PRCC or DeepChange or LaST
        ├── train
        ├── query
        ├── gallery

Training

# LTCC
python main.py --gpu_devices 0 --dataset ltcc --dataset_root DATASET_ROOT --dataset_filename LTCC-reID --save_dir SAVE_DIR --save_checkpoint

# Celeb-reID
python main.py --gpu_devices 0 --dataset celeb --dataset_root DATASET_ROOT --dataset_filename Celeb-reID --num_instances 4 --save_dir SAVE_DIR --save_checkpoint

# PRCC
python main.py --gpu_devices 0,1 --dataset prcc --dataset_root DATASET_ROOT --dataset_filename PRCC --max_epoch 30 --save_dir SAVE_DIR --save_checkpoint

# DeepChange
python main.py --gpu_devices 0,1 --dataset deepchange --dataset_root DATASET_ROOT --dataset_filename DeepChange --train_batch 64 --fg_start_epoch 45 --save_dir SAVE_DIR --save_checkpoint

# LaST
python main.py --gpu_devices 0,1 --dataset last --dataset_root DATASET_ROOT --dataset_filename LaST --train_batch 64 --num_instances 4 --fg_start_epoch 45 --save_dir SAVE_DIR --save_checkpoint

--dataset_root : replace DATASET_ROOT with your dataset root path

--save_dir: replace SAVE_DIR with the path to save log file and checkpoints

It is worth mentioning that adjusting the scanning radius of DBSCAN (by setting --eps) can explore fine-grained information of different granularities. Increasing the value of eps on difficult datasets (e.g., LTCC and DeepChange) may reduce noise and bring slightly better performance, while eps=0.4 mostly works well.

Evaluation

python main.py --gpu_devices 0 --dataset DATASET --dataset_root DATASET_ROOT --dataset_filename DATASET_FILENAME --resume RESUME_PATH --save_dir SAVE_DIR --evaluate

--dataset: replace DATASET with the dataset name

--dataset_filename: replace DATASET_FILENAME with the folder name of the dataset

--resume: replace RESUME_PATH with the path of the saved checkpoint

The above three arguments are set corresponding to Training.

Results

Celeb-reID

Backbone	Rank-1	Rank-5	mAP
ResNet-50	64.0	78.8	18.2

LTCC

Backbone	Setting	Rank-1	mAP
ResNet-50	Cloth-Changing	44.6	19.1
ResNet-50	Standard	75.9	39.9

PRCC

Backbone	Setting	Rank-1	mAP
ResNet-50	Cloth-Changing	65.0	63.1
ResNet-50	Standard	100	99.5

DeepChange

Backbone	Rank-1	mAP
ResNet-50	57.9	20.0

LaST

Backbone	Rank-1	mAP
ResNet-50	75.0	32.2

You can achieve similar results with the released code.

Citation

Please cite the following paper in your publications if it helps your research:

@article{wang2024exploring,
  title={Exploring fine-grained representation and recomposition for cloth-changing person re-identification},
  author={Wang, Qizao and Qian, Xuelin and Li, Bin and Xue, Xiangyang and Fu, Yanwei},
  journal={IEEE Transactions on Information Forensics and Security},
  volume={19},
  pages={6280-6292},
  year={2024},
  publisher={IEEE}
}

Contact

Any questions or discussions are welcome!

Qizao Wang (qzwang22@m.fudan.edu.cn)

QizaoWang / FIRe-CCReID

readme