Skip to content

Latest commit

 

History

History
149 lines (111 loc) · 5.45 KB

README.md

File metadata and controls

149 lines (111 loc) · 5.45 KB
Raw

PGAN

This code is for the paper Part-Guided Attention Learning for Vehicle Instance Retrieval. (TITS2020)

Note: this paper is also known as Part-Guided Attention Learning for Vehicle Re-Identification in Arxiv version.

[PDF] [Arxiv] [Project]

This code is ONLY released for academic use.

Model Structure

Installation

  • Python 3.6.5
  • Pytorch 1.0.0
  • Torchvision 0.2.1
  • scipy 1.2.0
  • pytorch-ignite 0.2.0
  • Please refer to requirements.txt for the other packages with the corresponding versions.

Preparation

  1. Run git clone https://github.com/zhangxinyu-xyz/PGAN-VehicleRe-ID-self.git

  2. Prepare dataset

    a. Download datasets: VeRi, PKU-VehicleID, VRIC, VERI-Wild. Move them to $PGAN/vehicle_data/.

    b. Download masks of part regions: VeRi, PKU-VehicleID, VRIC, VERI-Wild. Move them to $PGAN/vehicle_data/masks/.

    c. Decompress all datasets. Insure the data folder like the following structure (otherwise the data path should be modified in $PGAN/data/datasets/[DATANAME].py):

$PGAN/vehicle_data/
    VeRi/
        image_train
        image_query
        image_test
    PKU-VehicleID/
        image
        train_test_split
    VRIC/
        train_images
        probe_images
        gallery_images
    VERI-WILD/
        images
        train_test_split
    masks/
        VeRi/
            image_train
            image_query
            image_test
        PKU-VehicleID/
            image
        VRIC/
            train_images
            probe_images
            gallery_images
        VERI-WILD/
            images

  1. Download pre-trained model

    a. If you want to evaluate our method first, you may directly download the pre-trained models. The download links are: VeRi/model_best.pth, PKU-VehicleID/model_best.pth, VRIC/model_best.pth, VERI-WILD/model_best.pth.

    Note: I re-implement our methods and results may be slightly different from the original paper.

    b. Move them to $PGAN/PGAN_models/

Train

You can directly run Train_[DATANAME].sh file for the PGAN training process.

sh Train_veri.sh  ### train VeRi dataset
sh Train_vehicleid.sh  ### train PKU-VehicleID dataset
sh Train_vric.sh  ### train VRIC dataset
sh Train_veriwild.sh  ### train VERI-Wild dataset

If you want to train your own dataset, please add data file $PGAN/data/datasets/[DATANAME].py and generate masks of part regions.

Note that we use 2 GPUs to train VeRi and VRIC datasets, while 1 GPU to train PKU-VehicleID and VERI-Wild dataset.

Test

You can simply run Test_*.sh file for the evaluation of PGAN with pre-trained models.

sh Test_veri.sh  ### test VeRi dataset
sh Test_vehicleid.sh  ### test PKU-VehicleID dataset
sh Test_vric.sh  ### test VRIC dataset
sh Test_veriwild.sh  ### test VERI-Wild dataset

Results

Dataset folder mAP Top-1 Top-5
VeRi -- 79.4 96.3 98.6
PKU-VehicleID large 83.8 77.7 91.9
VRIC -- 85.1 78.1 93.5
VERI-Wild small 83.6 95.1 98.5
VERI-Wild medium 78.3 92.8 98.5
VERI-Wild large 70.6 89.2 95.7

Tips

You can modify SOLVER.ID_LOSS_WEIGHT to obtain better results for different datasets.

References

[1] Our code is conducted based on Strong ReID Baseline.

[2] Part-Guided Attention Learning for Vehicle Instance Retrieval, TITS2020

Citation

If you find this code useful in your research, please kindly consider citing our paper:

@article{zhang2020part,
  title={Part-guided attention learning for vehicle instance retrieval},
  author={Zhang, Xinyu and Zhang, Rufeng and Cao, Jiewei and Gong, Dong and You, Mingyu and Shen, Chunhua},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2020},
  publisher={IEEE}
}

or Arxiv version:

@article{zhang2019part,
  title={Part-guided attention learning for vehicle re-identification},
  author={Zhang, Xinyu and Zhang, Rufeng and Cao, Jiewei and Gong, Dong and You, Mingyu and Shen, Chunhua},
  journal={arXiv preprint arXiv:1909.06023},
  year={2019}
}

Contact

If you have any questions, please do not hesitate to contact us.

Xinyu Zhang

Rufeng Zhang

Chunhua Shen