Deep Meta Metric Learning (DMML)

This repo contains PyTorch code for ICCV19' paper: Deep Meta Metric Learning, including person re-identification experiments on Market-1501 and DukeMTMC-reID datasets.

Requirements

• Python 3.6+
• PyTorch 0.4
• tensorboardX 1.6

To install all python packages, please run the following command:

pip install -r requirements.txt

Datasets

Downloading

• Market-1501 dataset can be downloaded from here.
• DukeMTMC-reID dataset can be downloaded from here.

Preparation

After downloading the datasets above, move them to the datasets/ folder in the project root directory, and rename dataset folders to 'market1501' and 'duke' respectively. I.e., the datasets/ folder should be organized as:

|-- market1501
    |-- bounding_box_train
    |-- bounding_box_test
    |-- ...
|-- duke
    |-- bounding_box_train
    |-- bounding_box_test
    |-- ...

Usage

Training

After adding dataset directory in demo.sh, simply run the following command to train DMML on Market-1501:

bash demo.sh

Usage instructions of all training parameters can be found in config.py.

Evaluation

To evaluate the performance of a trained model, run

python eval.py

which will output Rank-1, Rank-5, Rank-10 and mAP scores.

Citation

Please use the citation provided below if it is useful to your research:

Guangyi Chen, Tianren Zhang, Jiwen Lu, and Jie Zhou, Deep Meta Metric Learning, ICCV, 2019.

@inproceedings{chen2019deep,
  title={Deep Meta Metric Learning},
  author={Chen, Guangyi and Zhang, Tianren and Lu, Jiwen and Zhou, Jie},
  booktitle={ICCV},
  year={2019}
}