Home

Awesome

MLCReID

The implementation for the Unsupervised Person Re-identification via Multi-label Classification, which is accepted by CVPR2020

Preparetion

Prerequisites

Data preparation

please refer to ECN to prepare dataset, the file structure is like

MLCReID/data
├── market
│   └── boundingbox_train
│   └── boundingbox_test
│   └── bounding_box_train_camstyle
│   └── query
├── duke
│   └── boundingbox_train
│   └── boundingbox_test
│   └── bounding_box_train_camstyle
│   └── query
└── msmt17
│   └── boundingbox_train
│   └── boundingbox_test
│   └── bounding_box_train_camstyle
│   └── query

Pretrained model

ResNet-50 pretrained on ImageNet is needed for model initialization, download and put it into models/imagenet dictionary (you can also omit this step by changing the code in resnet.py so that torchvision will automatically download it)

mkdir models
mkdir models/imagenet

The file tree should be

MLCReID/models
└── imagenet
    └── resnet50-19c8e357.pth

Logs and Output

The training logs and checkpoints are saved in output dictionary.

mkdir output

Train and Test

We utilize 1 GTX-2080TI GPU for model training, the hyper-parameters are set in configure files in experiments dictionary.

For example, training on Market-1501:

python tools/train.py --experiments experiments/market.yml --gpus 0

If you want to train model on DukeMTMC-reID or MSMT17, just replace the configure files.

python tools/train.py --experiments experiments/duke.yml --gpus 0

python tools/train.py --experiments experiments/msmt17.yml --gpus 0

If you want to test model, change the MODEL_FILE in configure file to the model path, then run following command:

python tools/test.py --experiments experiments/market.yml --gpus 0

You can also find these commands in scripts dictionary.

References