Pytorch implementation of Deep Face Super-Resolution with Iterative Collaboration between Attentive Recovery and Landmark Estimation (CVPR 2020) [arXiv][CVF]

If you find our work useful in your research, please consider citing:

@inproceedings{ma2020deep,
  title={Deep Face Super-Resolution with Iterative Collaboration between Attentive Recovery and Landmark Estimation},
  author={Ma, Cheng and Jiang, Zhenyu and Rao, Yongming and Lu, Jiwen and Zhou, Jie},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2020}
}

• Python 3 (Recommend to use Anaconda)
• PyTorch >= 1.1.0
• NVIDIA GPU + CUDA
• Python packages: pip install numpy opencv-python tqdm imageio pandas matplotlib tensorboardX

CelebA dataset can be downloaded here. Please download and unzip the img_celeba.7z file.

Helen dataset can be downloaded here. Please download and unzip the 5 parts of All images.

Testing sets for CelebA and Helen can be downloaded from Google Drive or Baidu Drive (extraction code: 6qhx).

Landmark annotations for CelebA and Helen can be downloaded in the annotations folder from Google Drive or Baidu Drive (extraction code: 6qhx).

The pretrained models can also be downloaded from the models folder in the above links. Then please place them in ./models.

To train a model:

cd code
python train.py -opt options/train/train_(DIC|DICGAN)_(CelebA|Helen).json

• The json file will be processed by options/options.py. Please refer to this for more details.

• Before running this code, please modify option files to your own configurations including:

  • proper dataroot_HR and dataroot_LR paths for the data loader
  • proper info_path for the annotations
  • whether to use a pretrained model and the path of the model
  • the path of the pretrained Feedback HourGlass model (for face landmark detection) if not using a pretrained model
  • the path of the feature extraction model (LightCNN_feature.pth) if training a GAN model
  • whether to resume training
  • saving frequency for models and states
  • other hyperparameters
  • loss function, etc.

• During training, you can use Tesorboard to monitor the losses with tensorboard --logdir tb_logger/NAME_OF_YOUR_EXPERIMENT

To generate SR images by a model:

cd code
python test.py -opt options/test/test_(DIC|DICGAN)_(CelebA|Helen).json

• Similar to training, the configurations can be modified in the json file. Please refer to this for more details.
• You can find your results in results/{test_name}/{dataset_name}. The PSNR and SSIM values will be stored in result.json while the average results will be recorded in average_result.txt
• We provide our DIC and DICGAN models used in our paper that can be downloaded in the models folder from Google Drive or Baidu Drive (extraction code: 6qhx). Then you can modify the directory of pretrained model and LR image sets in option files and run test.py for a quick test.

To evaluate the SR results by landmark detection:

python eval_landmark.py --info_path /path/to/landmark/annotations --data_root /path/to/result/images

• First you need to download the HG_68_CelebA.pth from the from Google Drive or Baidu Drive (extraction code: 6qhx) and put it into the ./models directory.
• Results will be in /path/to/result/images/landmark_result.json and the averaged results will be in landmark_average_result.txt.

The code is based on SRFBN and hourglass-facekeypoints-detection