• Introduction
• Features
• Datasets
• Configuration
• Training
• Evaluation

Person Re-Identification can be seen as an image retrieval problem. Given one query image in one camera, the objective is to find the images of the same person in other cameras.

This project is currently in beta release.

Datasets:

• Market-1501
• DukeMTMC (Coming Soon)
• CUHK03 (Coming Soon)
• MSMT17 (Coming Soon)
• VeRi (Coming Soon)

Models

• TransReID [Code]
• LightMBN [Code]
• PLR-OSNet [Code]

Market1501 (256x128 Image Size)

The Market1501 dataset contains 1501 different individuals (people) collected from 6 cameras in Tsinghua University.

Train Set: 751 different people, 12936 images Test Set: 750 different people, 3368 query images, 19732 testing images

Dataset Structure:

|__ Market-1501-v15.09.15/
    |__ bounding_box_test/      /* 19732 images for testing
    |__ bounding_box_train/     /* 12936 images for training
    |__ gt_bbox/                /* 25259 images (not use)
    |__ gt_query/               /* 3368 queries (not use)
    |__ query/                  /* 3368 query images (id is searched in "bounding_box_test/" folder)

Each image is named like 0000_c1s1_000151_01.jpg. For Market1501, the image name contains the identity label and camera id. Check the naming rule at here.

Create a configuration file in configs/ folder. Sample configuration can be found here. Then edit the fields you think if it is needed. This configuration file is needed for all of training and evaluation scripts.

$ python train.py --cfg configs/CONFIG_FILE_NAME.yaml

$ python val.py --cfg configs/CONFIG_FILE_NAME.yaml