Multi-Pseudo Regularized Label for Generated Data in Person Re-Identification

This repository contains the code for our TIP paper Multi-Pseudo Regularized Label for Generated Data in Person Re-Identification.

The code is mainly modified from Person-reID_GAN.

To run this code

1.Data Generation (GAN)

The first stage is to generate fake images by DCGAN. We use the DCGAN code at https://github.com/layumi/DCGAN-tensorflow.

You can also directly download our generated data (24000 generated images for Market1501) from Google_Drive

2.Semi-supervised Learning

This repository includes two baseline code and the dMpRL-II method in our paper.

Models	Reference
train_res_iden_baseline.m	ResNet50 baseline (only use real data)
train_res_iden_sMpRL.m	sMpRL label for generated images (combine real and generated data)
train_res_iden_MpRL2.m	dMpRL-I label for generated images (combine real and generated data)
train_res_iden_MpRL3.m	dMpRL-II label for generated images (combine real and generated data)

• We propose MpRL virtual labels for generated data. Three strategies are used to train the combination of real and generated data. We named the three strategies as sMpRL, dMpRL-I and dMpRL-II respectively in our paper. You can find more detailed code for our MpRL in:

sMpRL (Static MpRL)

dMpRL-I (Dynamic MpRL-I: Dynamically Update MpRL from scratch)

dMpRL-II (Dynamic MpRL-II: Dynamically Update MpRL from the intermediate point)

Compile Matconvnet

We use the Matconvnet package, you can just download this repos and run gpu_compile.m in Matlab to compile functions. The Matconvnet package already included in this repos. There is no need to download Matconvnet from the official website.

We use Cuda-8.0 and Cudnn-5.1, our code does not support cudnn version > 5.1. If you have any problem in compiling, first, try to check your cudnn version.

Dataset

We take Market1501 as an example in this repos. Download Market1501 Dataset

Training and Testing

1. Download the ResNet-50 model pretrained on Imagenet. Creat a folder named as result. Put it in the ./result dir.

2. Run the training code: Simply run the ./train.m, the trained model will be saved at ./result/Model_Name/

   This step including the data preparation and training (baseline, sMpRL, dMpRL-I and dMpRL-II).

   For real data preparation: code/prepare_data/prepare_data.m.

   For real + generated data preparation: code/prepare_data/prepare_gan_data.m (for dMpRL-I and dMpRL-II)

   For real + generated data preparation: code/prepare_data/prepare_sMpRL_label4data.m (for sMpRL)

3. Evaluation: Run ./test/test_gallery_query_crazy.m to extract feature of images in the gallery and query set. They will store in a .mat file in test. Then you can use it to do evaluation.

   Run ./evaluation/zzd_evaluation_res_faster.m to get the rank-1 accuracy and mAP

   If you find some bugs in our code, pls do not hesitate to email me: [email protected]

Citation

Please cite this paper in your publications if it helps your research:

@article{huang2019multi,
  title={Multi-Pseudo Regularized Label for Generated Data in Person Re-Identification},
  author={Huang, Yan and Xu, Jingsong and Wu, Qiang and Zheng, Zhedong and Zhang, Zhaoxiang and Zhang, Jian},
  journal={IEEE Transactions on Image Processing},
  volume={28},
  number={3},
  pages={1391--1403},
  year={2019},
  publisher={IEEE}
}