[Paper]

Yixiao Ge*, Zhuowan Li*, Haiyu Zhao, Guojun Yin, Shuai Yi, Xiaogang Wang, and Hongsheng Li
Neural Information Processing Systems (NIPS), 2018 (* equal contribution)

Pytorch implementation for our NIPS 2018 work. With the proposed siamese structure, we are able to learn identity-related and pose-unrelated representations.

• Baidu Pan links of datasets and pretrained models have been updated.

• Python 3
• Pytorch (We run the code under version 0.3.1, maybe lower versions also work.)

• Install dependencies (e.g., visdom and dominate). You can install all the dependencies by:

pip install scipy, pillow, torchvision, sklearn, h5py, dominate, visdom

• Clone this repo:

git clone https://github.com/yxgeee/FD-GAN
cd FD-GAN/

We conduct experiments on Market1501, DukeMTMC-reID, CUHK03 datasets. We need pose landmarks for each dataset during training, so we generate the pose files by Realtime Multi-Person Pose Estimation. And the raw datasets have been preprocessed by the code in open-reid. Download the prepared datasets following below steps:

• Create directories for datasets:

mkdir datasets
cd datasets/

• Download these datasets through the links below, and unzip them in the same root path.
  Market1501: [Google Drive] [Baidu Pan]
  DukeMTMC-reID: [Google Drive] [Baidu Pan]
  CUHK03: [Google Drive] [Baidu Pan]

As mentioned in the original paper, there are three stages for training our proposed framework.

We use a Siamese baseline structure based on ResNet-50. You can train the model with follow commands,

python baseline.py -b 256 -j 4 -d market1501 -a resnet50 --combine-trainval \
					--lr 0.01 --epochs 100 --step-size 40 --eval-step 5 \
					--logs-dir /path/to/save/checkpoints/

You can train it on specified GPUs by setting CUDA_VISIBLE_DEVICES, and change the dataset name [market1501|dukemtmc|cuhk03] after -d to train models on different datasets.
Or you can download the pretrained baseline model directly following the link below,

• Market1501_baseline_model: [Google Drive] [Baidu Pan]
• DukeMTMC_baseline_model: [Google Drive] [Baidu Pan]
• CUHK03_baseline_model: [Google Drive] [Baidu Pan]

And test them with follow commands,

python baseline.py -b 256 -d market1501 -a resnet50 --evaluate --resume /path/of/model_best.pth.tar

We need to pretain FD-GAN with the image encoder part (E in the original paper and net_E in the code) fixed first. You can train the model with follow commands,

python train.py --display-port 6006 --display-id 1 \
	--stage 1 -d market1501 --name /directory/name/of/saving/checkpoints/ \
	--pose-aug gauss -b 256 -j 4 --niter 50 --niter-decay 50 --lr 0.001 --save-step 10 \
	--lambda-recon 100.0 --lambda-veri 0.0 --lambda-sp 10.0 --smooth-label \
	--netE-pretrain /path/of/model_best.pth.tar

You can train it on specified GPUs by setting CUDA_VISIBLE_DEVICES. For main arguments,

• --display-port: display port of visdom, e.g., you can visualize the results by localhost:6006.
• --display-id: set 0 to disable visdom.
• --stage: set 1 for Stage II, and set 2 for stage III.
• --pose-aug: choose from [no|erase|gauss] to make augmentations on pose maps.
• --smooth-label: smooth the label of GANloss or not.

Other arguments can be viewed in options.py. Also you can directly download the models for stage II,

• Market1501_stageII_model: [Google Drive] [Baidu Pan]
• DukeMTMC_stageII_model: [Google Drive] [Baidu Pan]
• CUHK03_stageII_model: [Google Drive] [Baidu Pan]

There are four models in each directory for separate nets.

Notice: If you use visdom for visualization by setting --display-id 1, you need to open a new window and run the script python -m visdom.server -port=6006 before running the main program, where -port should be consistent with --display-port.

Finetune the whole framework by optimizing all parts. You can train the model with follow commands,

python train.py --display-port 6006 --display-id 1 \
	--stage 2 -d market1501 --name /directory/name/of/saving/checkpoints/ \
	--pose-aug gauss -b 256 -j 4 --niter 25 --niter-decay 25 --lr 0.0001 --save-step 10 --eval-step 5 \
	--lambda-recon 100.0 --lambda-veri 10.0 --lambda-sp 10.0 --smooth-label \
	--netE-pretrain /path/of/100_net_E.pth --netG-pretrain /path/of/100_net_G.pth \
	--netDi-pretrain /path/of/100_net_Di.pth --netDp-pretrain /path/of/100_net_Dp.pth

You can train it on specified GPUs by setting CUDA_VISIBLE_DEVICES.
We trained this model on a setting of batchsize 256. If you don't have such or better hardware, you may decrease the batchsize (the performance may also drop). Or you can directly download our final model,

• Market1501_stageIII_model: [Google Drive] [Baidu Pan]
• DukeMTMC_stageIII_model: [Google Drive] [Baidu Pan]
• CUHK03_stageIII_model: [Google Drive] [Baidu Pan]

And test best_net_E.pth by the same way as mentioned in Stage I.

• scripts for generate pose landmarks.
• generate specified images.

Please cite our paper if you find the code useful for your research.

@inproceedings{ge2018fdgan,
  title={FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification},
  author={Ge, Yixiao and Li, Zhuowan and Zhao, Haiyu and Yin, Guojun and Wang, Xiaogang and Li, Hongsheng},
  booktitle={Advances in Neural Information Processing Systems},
  year={2018}
}

Our code is inspired by pytorch-CycleGAN-and-pix2pix and open-reid.