Introduction

Code for our paper "Meta Pairwise Relationship Distillation for Unsupervised Person Re-identification" (improved version).

Prerequisites

• Necessary packages listed in requirements.txt

• Training Data

  (Market-1501, DukeMTMC-reID and MSMT-17. You can download these datasets from Zhong's repo)

  Unzip all datasets and ensure the file structure is as follow:

  MetaPRD/examples/data    
  │
  └───market1501 OR dukemtmc OR msmt17
       │   
       └───DukeMTMC-reID OR Market-1501-v15.09.15 OR MSMT17_V1
           │   
           └───bounding_box_train
           │   
           └───bounding_box_test
           | 
           └───query
  

Usage

on Market-1501 (w/ GCN)

python examples/train_mprd.py -b 64 -a resnet50 -d market1501 --iters 400 --momentum 0.1 --eps 0.4 --num-instances 16 --use-hard --use-gcn

on DukeMTMC-reID (w/ GCN)

python examples/train_mprd.py -b 64 -a resnet50 -d dukemtmcreid --iters 400 --momentum 0.1 --eps 0.6 --num-instances 16 --use-hard --use-gcn

on MSMT17 (w/ GCN)

python examples/train_mprd.py -b 64 -a resnet50 -d MSMT17 --iters 400 --momentum 0.1 --eps 0.7 --num-instances 16 --use-hard --use-gcn

If do not want to use the GCN to guide CNN training, you can optimize the CNN by follows

on Market-1501 (w/o GCN)

python examples/train_mprd.py -b 64 -a resnet50 -d market1501 --iters 400 --momentum 0.1 --eps 0.4 --num-instances 16 --use-hard

on DukeMTMC-reID (w/o GCN)

python examples/train_mprd.py -b 64 -a resnet50 -d dukemtmcreid --iters 400 --momentum 0.1 --eps 0.6 --num-instances 16 --use-hard

on MSMT17 (w/o GCN)

python examples/train_mprd.py -b 64 -a resnet50 -d MSMT17 --iters 400 --momentum 0.1 --eps 0.7 --num-instances 16 --use-hard

pre-trained model

for CNN

When training with the backbone of IBN-ResNet, you need to download the ImageNet-pretrained model from this link The pre-trained model of CNN are saved in examples/pretrained ImageNet-pretrained models for ResNet-50 will be automatically downloaded in the python script. use -a resnet50 (default) for the backbone of ResNet-50, and -a resnet_ibn50a for the backbone of IBN-ResNet.

for GCN

If you want to use GCN to guide CNN training, we suggest utilizing the pre-trained model for GCN in mprd/gcn/saved_model/pretrained, more details in there.

Citation

If this repo is helpful for your research, please consider citing the paper:

@inproceedings{ji2021meta,
  title={Meta Pairwise Relationship Distillation for Unsupervised Person Re-identification},
  author={Ji, Haoxuanye and Wang, Le and Zhou, Sanping and Tang, Wei and Zheng, Nanning and Hua, Gang},
  booktitle={ICCV},
  pages={3661--3670},
  year={2021}
}

or

@article{ji2022meta,
  title={Meta Pairwise Relationship Distillation for Unsupervised Person Re-identification},
  author={Ji, Haoxuanye and Wang, Le and Zhou, Sanping and Tang, Wei and Zheng, Nanning and Hua, Gang},
  booktitle={submitted to T-NNLS},
  year={2022}
}

Acknowledgments

This repo borrows partially from Link-Prediction-Based-on-Graph-Neural-Networks, SpCL and cluster-contrast-reid. If you find our code useful, please cite their papers.

@inproceedings{nips2018link,
  title={Link prediction based on graph neural networks},
  author={Zhang, Muhan and Chen, Yixin},
  booktitle={NeurIPS},
  pages={5171--5181},
  year={2018}
}

@inproceedings{ge2020selfpaced,
    title={Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID},
    author={Yixiao Ge and Feng Zhu and Dapeng Chen and Rui Zhao and Hongsheng Li},
    booktitle={NeurIPS},
    year={2020}
}

@inproceedings{arxiv2021Cluster,
    author = {Dai, Zuozhuo and Wang, Guangyuan and Yuan, Weihao and Zhu, Siyu and Tan, Ping},
    title = {Cluster Contrast for Unsupervised Person Re-Identification},
    booktitle = {arXiv:2103.11568},
    year = 2021
}