Awesome

PyTorch implementation for Robust Object Re-identification with Coupled Noisy Labels (IJCV 2024).

LCNL extends the previous work Learning with Twin Noisy Labels for Visible-Infrared Person Re-Identification (CVPR 2022) by generalizing DART to both single- and cross-modality ReID tasks with improved loss function.

Introduction

LCNL framework

Requirements

• Python 3.7
• PyTorch ~1.7.1
• numpy
• scikit-learn

Datasets

SYSU-MM01 and RegDB

We follow ADP to obtain the datasets.

Market1501, Duke-MTMC and VeRi-776

We follow TransReID to obtain the datasets.

Visible-infrared ReID

Training

Modify the data_path and specify the noise_ratio to train the model.

# SYSU-MM01: noise_ratio = {0, 0.2, 0.5}
python run.py --gpu 0 --dataset sysu --data-path data_path --noise-rate 0. --savename sysu_lcnl_nr0 --op-type weighty

# RegDB: noise_ratio = {0, 0.2, 0.5}, trial = 1-10
python run.py --gpu 0 --dataset regdb --data-path data_path --noise-rate 0. --savename regdb_lcnl_nr0 --trial 1

Evaluation

Modify the data_path and model_path to evaluate the trained model.

# SYSU-MM01: mode = {all, indoor}
python test.py --gpu 0 --dataset sysu --data-path data-path --model_path model_path --resume-net1 'sysu_lcnl_nr0_net1.t' --resume-net2 'sysu_lcnl_nr0_net1.t' --mode all

# RegDB: --tvsearch or not (whether thermal to visible search)
python test.py --gpu 0 --dataset regdb --data-path data-path --model_path model_path --resume-net1 'regdb_lcnl_nr20_trial{}_net1.t' --resume-net2 'regdb_lcnl_nr20_trial{}_net2.t'

Reproduce

We provide the checkpoints (for evaluation) and noise indexes (should be placed on the dataset path for training) for result reproducing.

Todo

• Release the code for vehicle ReID task.
• Release the code for visible ReID task.

Citation

If LCNL is useful for your research, please consider citing:

@article{yang2024lcnl,
  title={Robust Object Re-identification with Coupled Noisy Labels},
  author={Yang, Mouxing and Huang, Zhenyu and Peng, Xi},
  journal={International Journal of Computer Vision},
  year={2024},
  publisher={Springer}
}

or the previous conference version:

@InProceedings{Yang_2022_CVPR,
    author={Yang, Mouxing and Huang, Zhenyu and Hu, Peng and Li, Taihao and Lv, Jiancheng and Peng, Xi},
    title={Learning With Twin Noisy Labels for Visible-Infrared Person Re-Identification},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month={June},
    year={2022},
    pages={14308-14317}
}

License

Apache License 2.0

Acknowledgements

The code is based on ADP, TransReID and DART licensed under Apache 2.0.