Multiple Granularity Network

Implement of paper:Learning Discriminative Features with Multiple Granularities for Person Re-Identification

Dependencies

Python >= 3.5
PyTorch >= 0.4.0
torchvision
scipy
numpy
scikit_learn

Current Result

Re-Ranking	backbone	mAP	rank1	rank3	rank5	rank10
yes	resnet50	94.33	95.58	97.54	97.92	98.46
no	resnet50	86.15	94.95	97.42	98.07	98.93

Data

The data structure would look like:

data/
    bounding_box_train/
    bounding_box_test/
    query/

Market1501

Download from here

DukeMTMC-reID

Download from here

CUHK03

Download cuhk03 dataset from "http://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html"
Unzip the file and you will get the cuhk03_release dir include cuhk-03.mat
Download "cuhk03_new_protocol_config_detected.mat" from "https://github.com/zhunzhong07/person-re-ranking/tree/master/evaluation/data/CUHK03" and put it with cuhk-03.mat. We need this new protocol to split the dataset.

python utils/transform_cuhk03.py --src <path/to/cuhk03_release> --dst <path/to/save>

NOTICE:You need to change num_classes in network depend on how many people in your train dataset! e.g. 751 in Market1501

Weights

Pretrained weight download from google drive

Train

You can specify more parameters in opt.py

python main.py --mode train --data_path <path/to/Market-1501-v15.09.15>

Evaluate

Use pretrained weight or your trained weight

python main.py --mode evaluate --data_path <path/to/Market-1501-v15.09.15> --weight <path/to/weight_name.pt>

Visualize

Visualize rank10 query result of one image(query from bounding_box_test)

Extract features will take a few munutes, or you can save features as .mat file for multiple uses

python main.py --mode vis --query_image <path/to/query_image> --weight <path/to/weight_name.pt>

Citation

@ARTICLE{2018arXiv180401438W,
    author = {{Wang}, G. and {Yuan}, Y. and {Chen}, X. and {Li}, J. and {Zhou}, X.},
    title = "{Learning Discriminative Features with Multiple Granularities for Person Re-Identification}",
    journal = {ArXiv e-prints},
    year = 2018,
}