Leader-based Multi-Scale Attention Deep Architecture for Person Re-identification

This repository contains the Pytorch implementation for the paper "Leader-based Multi-Scale Attention Deep Architecture for Person Re-identification"

Framework

Getting Started

Prerequisites

• Python 3.6 or 3.7
• Pytorch >= 1.1.0
• tensorboardX

Prepare data

Please download Market-1501 dataset and organize it as follows

MuDeep_v2
    ├── dataset
    │      └─ Market-1501 # for Market-1501 dataset
    │             ├── bounding_box_train
    │             ├── bounding_box_test
    │             ├── query
    │
    ├── train.py

How to train

In config.py, set configurations for training, including NAME, GPU_ID and ROOT. You can keep others as default to reproduce the result.

# example
__C.NAME = 'market'  # name your model, the model files (.pkl) will be saved according to this name
__C.GPU_ID = 0,1  
__C.ROOT = '/home/qxl/work/mudeep_v2/'  # path to your project folder, all models and log files will be saved in this folder
...

In train.py, using the following command lines to train the model

# example
engine = MuDeep_v2(cfg)
engine.train()

Once trained, the models and log file will be saved in ROOT/model/NAME/ and ROOT/log/NAME/ respectively.

By default, we evaluate the model every 5 epochs, the results will be written in ROOT/model/NAME/opt.txt

How to evaluate

In train.py, using the following command lines to evaluate the model

# example
engine = MuDeep_v2(cfg)
engine.test(model_path='/home/qxl/work/mudeep_v2/model/market',   # path to your model
            out_name='market_evaluate'  # name the output TXT file
           )

Result

Supervised Learning

Name	Backbone	image size	mAP	Rank-1	Rank-5	Rank-10	url
market_v1	ResNet-50	384 x 192	85.61	95.46	98.43	98.93	download
market_v2	ResNet-50	384 x 128	85.39	95.25	98.22	98.99	download

Domain Generalization

Name	Backbone	image size	Rank-1/mAP
			DukeMTMC-reID	CUHK03-np Detected	CUHK03-np Labeled
market_v1	ResNet-50	384 x 192	46.68/27.33	10.64/8.36	11.79/9.34
market_v2	ResNet-50	384 x 128	48.56/28.24	12.00/10.08	13.00/10.70

Citation

If you find this project useful in your research, please consider cite:

@article{qian2019leader,
  title={Leader-based multi-scale attention deep architecture for person re-identification},
  author={Qian, Xuelin and Fu, Yanwei and Xiang, Tao and Jiang, Yu-Gang and Xue, Xiangyang},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  volume={42},
  number={2},
  pages={371--385},
  year={2019},
  publisher={IEEE}
}

Contact

Any questions or discussion are welcome!

Xuelin Qian ([email protected])