xlearning-scu/2022-cvpr-dart

PyTorch implementation for Learning with Twin Noisy Labels for Visible-Infrared Person Re-Identification (CVPR 2022).

PyTorch implementation for Learning with Twin Noisy Labels for Visible-Infrared Person Re-Identification (CVPR 2022).

Introduction

DART framework

Requirements

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

Datasets

SYSU-MM01 and RegDB

We follow ADP to obtain datasets.

Training and Evaluation

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 0.2 --savename sysu_dart_nr20 

# RegDB: noise_ratio = {0, 0.2, 0.5}, trial = 1-10
python run.py --gpu 0 --dataset regdb --data-path data_path --noise-rate 0.2 --savename regdb_dart_nr20 --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_dart_nr20_net1.t' --resume-net2 'sysu_dart_nr20_net2.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_dart_nr20_trial{}_net1.t' --resume-net2 'regdb_dart_nr20_trial{}_net2.t'

Citation

If DART is useful for your research, please cite the following paper:

@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 licensed under Apache 2.0.

2022-cvpr-dart's Issues

Link of the paper

The original link is wrong.

The paper link should be: http://pengxi.me/wp-content/uploads/2022/03/Learning-with-Twin-Noisy-Labels-for-Visible-Infrared-Person-Re-Identification.pdf

出了一个bug，请问是什么原因。

==> Preparing Data Loader...
Traceback (most recent call last):
File "/home/2022-CVPR-DART/run.py", line 504, in
prob_A_V, prob_A_I = eval_train(net1, eval_loader, 'A')
File "/home/2022-CVPR-DART/run.py", line 313, in eval_train
losses_V_aug2[index_V[n1 + 32]] = loss1[n1 + 32]
IndexError: index 32 is out of bounds for axis 0 with size 32

data_path路径问题

请问在哪个文件下面修改data_path呢

Acquisition of the noise id

Hello, I have some doubts about the part of the code that differentiates TN, and whether the use of rgb_noiseIdx for differentiation is causing data leakage. By definition, the model should not be able to directly obtain which noise ids are included. I'm not sure if I'm wrong about this variable. Can u give me an answer?
T = (prob_A_V[evaltrainset.rgb_noiseIdx] < args.p_threshold). sum() / len(evaltrainset.rgb_noiseIdx)

可视化

我想知道是否有TP-FP概率分布的代码