CLIP-ReID: Exploiting Vision-Language Model for Image Re-Identification without Concrete Text Labels [pdf]

Pipeline

Installation

conda create -n clipreid python=3.12.2
conda activate clipreid
pip install -r requirements.txt

Prepare Dataset

Download the datasets (Market-1501, MSMT17, DukeMTMC-reID, Occluded-Duke, VehicleID, VeRi-776), and then unzip them to your_dataset_dir.

Training

For example, if you want to run CNN-based CLIP-ReID-baseline for the Market-1501, you need to modify the bottom of configs/person/cnn_base.yml to

DATASETS:
   NAMES: ('market1501')
   ROOT_DIR: ('your_dataset_dir')
OUTPUT_DIR: 'your_output_dir'

then run

CUDA_VISIBLE_DEVICES=0 python train.py --config_file configs/person/cnn_base.yml

if you want to run ViT-based CLIP-ReID for MSMT17, you need to modify the bottom of configs/person/vit_clipreid.yml to

DATASETS:
   NAMES: ('msmt17')
   ROOT_DIR: ('your_dataset_dir')
OUTPUT_DIR: 'your_output_dir'

then run

CUDA_VISIBLE_DEVICES=0 python train_clipreid.py --config_file configs/person/vit_clipreid.yml

if you want to run ViT-based CLIP-ReID+SIE+OLP for MSMT17, run:

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]'

Evaluation

For example, if you want to test ViT-based CLIP-ReID for MSMT17

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'

Acknowledgement

Codebase from TransReID, CLIP, and CoOp.

The veri776 viewpoint label is from https://github.com/Zhongdao/VehicleReIDKeyPointData.

Trained models and test logs

Datasets	MSMT17	Market	Duke	Occ-Duke	VeRi	VehicleID
CNN-baseline	model|test	model|test	model|test	model|test	model|test	model|test
CNN-CLIP-ReID	model|test	model|test	model|test	model|test	model|test	model|test
ViT-baseline	model|test	model|test	model|test	model|test	model|test	model|test
ViT-CLIP-ReID	model|test	model|test	model|test	model|test	model|test	model|test
ViT-CLIP-ReID-SIE-OLP	model|test	model|test	model|test	model|test	model|test	model|test

Note that all results listed above are without re-ranking.

With re-ranking, ViT-CLIP-ReID-SIE-OLP achieves 86.7% mAP and 91.1% R1 on MSMT17.

Citation

If you use this code for your research, please cite

@article{li2022clip,
  title={CLIP-ReID: Exploiting Vision-Language Model for Image Re-Identification without Concrete Text Labels},
  author={Li, Siyuan and Sun, Li and Li, Qingli},
  journal={arXiv preprint arXiv:2211.13977},
  year={2022}
}