/reid_CAN

The 3st place solution for PRCV2020.

Primary LanguagePythonMIT LicenseMIT

CAN: Collaborative Attention Network for Person Re-identification

In this repository, we include the 3th place solution for LSPRC at PRCV 2020.Our paper

Authors

Dependencies

  • Python == 3.6.2
  • PyTorch == 1.3.0
  • TorchVision == 0.4.0
  • Matplotlib == 3.0.2
  • Sklearn == 0.20.1
  • PIL == 5.3.0
  • Scipy == 1.1.0
  • Tqdm == 4.28.1
  • prefetch_generator = 1.0.1
  • OpenCV == 3.4.3

Warning

The batchnorm.py file from PyTorch is modified in our model which is under the path of './model'. Please ensure that the path of modules and nn is consistent with the system environment. More details please refer to the line5 and line10 in batchnorm.py.

Data

The data structure would look like:

data/
    bounding_box_train/
    bounding_box_test/
    query/

  • Market1501 Download from here

  • DukeMTMC-reID Download from here

Project File Structure

|--reid_CAN
  |--data               #（Train data related）
    |--_init_.py
    |--commom.py
    |--market1501.py
    |--sampler.py
  |--loss               #（Loss function）
    |--_init_.py
    |--center_loss.py
    |--triplet_loss.py
  |--model              #（Model structure）
    |--backbones
      |--ibnnet
        |--__init__.py
        |--modules.py
        |--resnet_ibn.py
      |--__init__.py
    |--pretrained
      |--resnet50_ibn_b-9ca61e85.pth
    |--_init_.py
    |--batchnorm.py
    |--can.py
    |--layers.py
  |--utils		#（Functions）
    |--__init__.py
    |--Adam.py
    |--functions.py
    |--functions_camera.py
    |--lr_scheduler.py
    |--random_erasing.py
    |--random_patch.py
    |--utlity.py
  |--main.py
  |--option.py          #（Parameter details）
  |--README.md
  |--run.sh
  |--trainer.py

Get Started

cd ./reid_CAN
sh run.sh

In the run.sh file, cancel the comment of the training or testing command. and run sh run.sh, you can quick stark with training or testing.

Train

You can specify more parameters in option.py. Change the parameters --datadir to your own train data path

CUDA_VISIBLE_DEVICES=0 python3 main.py --datadir ./../REID_datasets/prcv2020_v2 --backbone_name resnet50_ibn_b --batchid 8 --batchimage 4 --batchtest 16 --test_every 10 --epochs 601 --loss 1*CrossEntropy_Loss+1*Triplet_Loss+0.0005*Center_Loss --margin 1.3 --nGPU 1 --lr 3.5e-4 --optimizer ADAM_GCC --reset --amsgrad --num_classes 1295 --height 384 --width 192 --save_models --save prcv2020_train_v2-can

Parameter details are as follows：

  • -CUDA_VISIBLE_DEVICES: GPU ID for training.
  • -batchid: Number of training IDs sent to each batch
  • -batchimage: Number of images sent to training per ID
  • -test_every: How many epochs do a test
  • -epochs: Total training epoch number
  • -loss: Weights * loss function
  • -margin: Tripletloss parameter
  • -nGPU: How many GPU choose to train
  • -lr: Learning rate
  • -num_classes: Number of independent IDs
  • -save_models: Save model for each test
  • -save: Model saved path

Refer to option.py for other related parameters

Test

Change the parameters --datadir to your own test data path

CUDA_VISIBLE_DEVICES=0 python3 main.py --datadir ./../REID_datasets/prcv2020_v2 --backbone_name resnet50_ibn_a --margin 1.3 --nGPU 1 --test_only --resume 0 --pre_train model_best.pt --batchtest 16 --num_classes 1295 --height 384 --width 192 --save test-prcv2020_v2_can

Parameter details are as follows：

  • -test_only: Test switch, only test if open
  • -pre_train: The model choosed to test

Refer to option.py for other related parameters

Current Result

  • Market-1501
Re-Ranking mAP Rank-1 Rank-3 Rank-5 Rank-10
w/o 90.69 96.32 98.04 98.72 99.35
with 95.41 96.56 97.95 98.34 98.84
  • DukeMTMC-ReID
Re-Ranking mAP Rank-1 Rank-3 Rank-5 Rank-10
w/o 82.27 91.56 95.02 96.01 97.17
with 91.85 93.54 95.78 96.41 96.99
  • CUHK03
Re-Ranking mAP Rank-1 Rank-3 Rank-5 Rank-10
w/o 82.82 86.14 91.36 93.71 96.00
with 91.92 90.93 92.29 94.07 97.07

Citation

If you find our work useful in your research, please consider citing:

@article{li2021collaborative,
    title={Collaborative Attention Network for Person Re-identification},
    author={Li, Wenpeng and Sun, Yongli and Wang, Jinjun and Cao, Junliang and Xu, Han and Yang, Xiangru and Sun, Guangze and Ma, Yangyang and Long, Yilin},
    journal={Journal of Physics: Conference Series},
    doi={10.1088/1742-6596/1848/1/012074},
    year={2021},
}