This repo contains the official codes for our paper:

Exploit Camera Raw Data for Video Super-Resolution via Hidden Markov Model Inference
Xiaohong Liu, Kangdi Shi, Zhe Wang, Jun Chen

Dependencies and Installation

  1. Clone repo

    $ git clone https://github.com/proteus1991/RawVSR.git
  2. Install dependent packages

    $ cd RawVSR
    $ pip install -r requirements.txt
  3. Setup the Deformable Convolution Network (DCN)

    Since our RawVSR use the DCN for feature alignment extracted from different video frames, we follow the setup in EDVR, where more details can be found.

    $ python setup.py develop

    Note that the deform_conv_cuda.cpp and deform_conv_cuda_kernel.cu have been modified to solve compile errors in PyTorch >= 1.7.0. If your PyTorch version < 1.7.0, you may need to download the original setup code.


  • train.py and test.py are the entry codes for training and testing the RawVSR.
  • ./data/ contains the codes for data loading.
  • ./dataset/ contains the corresponding video sequences.
  • ./dcn/ is the dependencies of DCN.
  • ./models/ contains the codes to define the network.
  • ./utils/ includes the utilities.
  • ./weight_checkpoint/ saves checkpoints and the best network weight.

Raw Video Dataset (RawVD)

Since we are not aware of the existence of publicly available raw video datasets, to train our RawVSR,
a raw video dataset dubbled as RawVD is built.

In this dataset, we provide the ground-truth sRGB frames in folder 1080p_gt_rgb. Low-resolution (LR) Raw frames are in folder 1080p_lr_d_raw_2 and 1080p_lr_d_raw_4 in terms of different scale ratios.
Their corresponding sRGB frames are in folder 1080p_lr_d_rgb_2 and 1080p_lr_d_rgb_4, where d in folder name stands for the degradations including defocus blurring and heteroscedastic Gaussian noise.
We also released the original raw videos in Magic Lantern Video (MLV) format. The corresponding software to play it can be found here.
Details can be found in Section 3 of our paper.

Quick Start

1. Testing

Make sure all dependencies are successfully installed.

Run test.py with --scale_ratio and save_image tags.

$ python test.py --scale_ratio 4 --save_image

The help of --scale_ratio and save_image tags is shown by running:

$ python test.py -h

If everything goes well, the following messages will appear in your bash:

--- Hyper-parameter default settings ---
train settings:
 {'dataroot_GT': '/media/lxh/SSD_DATA/raw_test/gt/1080p/1080p_gt_rgb', 'dataroot_LQ': '/media/lxh/SSD_DATA/raw_test/w_d/1080p/1080p_lr_d_raw_4', 'lr': 0.0002, 'num_epochs': 100, 'N_frames': 7, 'n_workers': 12, 'batch_size': 24, 'GT_size': 256, 'LQ_size': 64, 'scale': 4, 'phase': 'train'}
val settings:
 {'dataroot_GT': '/media/lxh/SSD_DATA/raw_test/gt/1080p/1080p_gt_rgb', 'dataroot_LQ': '/media/lxh/SSD_DATA/raw_test/w_d/1080p/1080p_lr_d_raw_4', 'N_frames': 7, 'n_workers': 12, 'batch_size': 2, 'phase': 'val', 'save_image': True}
network settings:
 {'nf': 64, 'nframes': 7, 'groups': 8, 'back_RBs': 4}
dataset settings:
 {'dataset_name': 'RawVD'}
--- testing results ---
store: 29.04dB
painting: 29.02dB
train: 28.59dB
city: 29.08dB
tree: 28.06dB
avg_psnr: 28.76dB
--- end ---

The RawVSR is tested on our elaborately-collected RawVD. Here the PSNR results should be the same as Table 1 in our paper.

2. Training

Run train.py without --save_image tag to reduce the training time.

$ python train.py --scale_ratio 4

If you want to change the default hyper-parameters (e.g., modifying the batch_size), simply go config.py. All network and training/testing settings are stored there.


Some codes (e.g., DCN) are borrowed from EDVR with modification.


If you use this code, please kindly cite

  title={Exploit Camera Raw Data for Video Super-Resolution via Hidden Markov Model Inference},
  author={Liu, Xiaohong and Shi, Kangdi and Wang, Zhe and Chen, Jun},
  journal={arXiv preprint arXiv:2008.10710},


Should you have any question about this code, please open a new issue directly. For any other questions, you might contact me in email: [email protected].