shuxjweb / last Goto Github PK

[Project] [Paper]

This repository contains the source code for loading the LaST dataset and evaluating its generalization.

LaST is a large-scale dataset with more than 228k pedestrian images. It is used to study the scenario that pedestrians have a large activity scope and time span. Although collected from movies, we have selected suitable frames and labeled them as carefully as possible. Besides the identity label, we also labeled the clothes of pedestrians in the training set.

• Train: 5000 identities and 71,248 images.
• Val: 56 identities and 21,379 images.
• Test: 5806 identities and 135,529 images.

Note: You can download LaST from these links: BaiduPan with password: vvfe. or Googledrive.

• Python 3.7
• PyTorch 1.6
• Torchvision 0.7.0
• Cuda10.2

python last_train_bot.py --train 1 --data_dir /data/last/ --logs_dir ./20210407_last_bot_base

1. Put the pre-trained model in the folder "pre_feat". For example, last_ini_imagenet.pth.

./pre_feat/last_ini_imagenet.pth

2. Modify the loaded model name as follows:

last_model_wts = torch.load(os.path.join('pre_feat', 'last_ini_imagenet.pth'))

3. Start Testing

python prcc_train_base_last.py --train 0 --data_dir /data/prcc/ --logs_dir ./pre_feat

1. Put the pre-trained model in the folder "pre_feat". For example, last_ini_imagenet.pth.

./pre_feat/last_ini_imagenet.pth

2. Start Training

python prcc_train_base_last.py --train 1 --data_dir /data/prcc/ --logs_dir ./20210205_prcc_base_last_sgd

Please kindly cite this paper in your publications if it helps your research:

@article{shu2021large,
  title={Large-Scale Spatio-Temporal Person Re-identification: Algorithm and Benchmark},
  author={Shu, Xiujun and Wang, Xiao and Zang, Xianghao and Zhang, Shiliang and Chen, Yuanqi and Li, Ge and Tian, Qi},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  pages={4390-4403},
  year={2022}
}

We forked the projects in Person_reID_baseline_pytorch, fast-reid, deep-person-reid and reid-strong-baseline. Thank the authors for their great work.

The dataset and code are released for academic research use only. If you have questions, please contact [email protected]