A Tiny Person ReID Baseline

Paper: "Bag of Tricks and A Strong Baseline for Deep Person Re-identification"[pdf]

This project refers the official code link and can reproduce the results as good as it on Market1501 when the input size is set to 256x128. If you find this project useful, please cite the offical paper.

@inproceedings{luo2019bag,
  title={Bag of Tricks and A Strong Baseline for Deep Person Re-identification},
  author={Luo, Hao and Gu, Youzhi and Liao, Xingyu and Lai, Shenqi and Jiang, Wei},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  year={2019}
}

Updates (Difference from Official Code)

v0.1.1 (Sep. 2019)
- Support ArcFace loss, which can converge faster (around 50 epochs) and achieve slightly better performance than softmax+triplet loss+center loss
- Support visualizing reID results
- Add comments in config.py
v0.1.0 (Jun. 2019)
- Develop based on the pytorch template link
- No need to install ignite and yacs
- Support computing DistMap using cosine similarity
- Set hyperparameters using a configuration class
- Only support ResNet50 as the backbone

Directory layout

.
├── config                  # hyperparameters settings
│   └── ...                 
├── datasets                # dataloader
│   └── ...           
├── log                     # log and model weights             
├── loss                    # loss function code
│   └── ...   
├── model                   # model
│   └── ...  
├── processor               # training and testing procedures
│   └── ...    
├── solver                  # optimization code
│   └── ...   
├── utils                   # metrics code
│   └── ...   
├── train.py                # train code 
├── test.py                 # test code 
├── get_vis_result.py       # get visualized results 
├── imgs                    # images for readme              
└── README.md

Pipeline

Results on Market1501 (rank1/rank5/rank10/mAP)

Model	Loss	Market1501
ResNet50 (128x64)	softmax+triplet+center	88.2/95.7/97.5/70.5
ResNet50 (256x128)	softmax+triplet+center	94.0/96.9/98.1/83.4
ResNet50 (256x128)	arcface	94.7/97.7/98.3/84.3

The pretrained (128x64) model can be downloaded now. Extraction code is u3q5.

Get Started

cd to folder where you want to download this repo
Run git clone https://github.com/lulujianjie/person-reid-tiny-baseline.git
Install dependencies:
- pytorch>=0.4
- torchvision
- cv2 (optional)

Train

python train.py

Test

python test.py

To get visualized reID results, first create results folder in log dir, then:

python get_vis_result.py

You will get the ranked results (query|rank1|rank2|...), like:

FelixZhang7/person-reid-tiny-baseline