Graph-based Person Signature for Person Re-Identifications (GPS)

This repository is the implementation of GPS for Person Re-Identifications task. Our model achieved 87.8, 78.7 on mean Average Precision (mAP) and 95.2, 88.2 on Cumulative Matching Characteristic (CMC) R-1 over Market1501 and DukeMTMC-ReID datasets, respectively. For the detail, please refer to link.

This repository is based on and inspired by @Hao Luo's work. We sincerely thank for their sharing of the codes.

Summary

• The proposed framework
• Prerequisites
• Datasets
• Training
• Testing
• Citation
• License
• More information

The proposed framework

Prerequisites

Python3

Please install dependence package by run following command:

pip install -r requirements.txt

Datasets

Market1501

• The Market1501 original dataset can be downloaded via link.

• The Market1501 attributes and body-part masks can be downloaded via link.

• Additionally, a full version including all the above data is available here. Downloading this version is sufficient.

• Extract the downloaded files to the dataset/market1501_gps_release/ directory.

This directory is constructed as follow:

|---dataset   
|---|---market1501_gps_release   
|---|---|---bounding_box_test
|---|---|---bounding_box_train
|---|---|---gt_bbox
|---|---|---gt_query
|---|---|---query
|---|---|---attribute
|---|---|---Masks
|---|---|---adj.pkl
|---|---|---glove.pkl
|---|---|---image_mask_path_dict.pkl
|---|---|...

DukeMTMC-ReID

• The DukeMTMC-ReID original dataset can be downloaded via link.

• The DukeMTMC-ReID attributes and body-part masks can be downloaded via link.

• Additionally, a full version including all the above data is available here. Downloading this version is sufficient.

• Extract the downloaded files to the dataset/dukemtmc_gps_release/ directory.

This directory is constructed as follow:

|---dataset   
|---|---dukemtmc_gps_release
|---|---|---bounding_box_test
|---|---|---bounding_box_train
|---|---|---query
|---|---|---attribute
|---|---|---Masks
|---|---|---adj.pkl
|---|---|---glove.pkl
|---|---|---image_mask_path_dict.pkl
|---|---|...

Thanks to Yutian Lin (github) for providing the Market1501 and DukeMTMC-ReID attributes.

Training

You should download the pretrained weight of ResNet50 model via link and put to pretrained/resnet50-pretrained/ directory.

To train GPS model on Market1501 dataset, please follow:

$ python train.py --config_file configs/market1501_gps_softmax_triplet_center.yml

To train GPS model on DukeMTMC-ReID dataset, please follow:

$ python train.py --config_file configs/dukemtmc_gps_softmax_triplet_center.yml

The training scores will be printed every epoch.

Testing

In this repo, we include the pre-trained weight of GPS_market1501 and GPS_dukemtmc models.

For GPS_market1501 pretrained model. Please download the link and move to pretrained/ directory. The trained GPS_market1501 model can be tested in Market1501 test split via:

$ python test.py --config_file configs/market1501_gps_softmax_triplet_center.yml MODEL.PRETRAIN_CHOICE "('self')" TEST.WEIGHT "('pretrained/GPS_market1501.pth')"

For GPS_dukemtmc pretrained model. Please download the link and move to pretrained. The trained GPS_dukemtmc model can be tested in DukeMTMC-ReID test split via:

$ python test.py --config_file configs/dukemtmc_gps_softmax_triplet_center.yml MODEL.PRETRAIN_CHOICE "('self')" TEST.WEIGHT "('pretrained/GPS_dukemtmc.pth')"

Citation

If you use this code as part of any published research, we'd really appreciate it if you could cite the following paper:

@InProceedings{Nguyen_2021_CVPR,
    author    = {Nguyen, Binh X. and Nguyen, Binh D. and Do, Tuong and Tjiputra, Erman and Tran, Quang D. and Nguyen, Anh},
    title     = {Graph-Based Person Signature for Person Re-Identifications},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {3492-3501}
}

License

MIT License

AIOZ © 2021 All rights reserved.

More information

AIOZ AI Homepage: https://ai.aioz.io

AIOZ Network: https://aioz.network