Deformable 3D Convolution for Video Super-Resolution

Pytorch implementation of deformable 3D convolution network (D3Dnet). [PDF]

Our code is based on cuda and can perform deformation in any dimension of 3D convolution.

Overview

Architecture of D3Dnet

Architecture of D3D

Requirements

Python 3
pytorch (1.2.0), torchvision (0.4.0)
numpy, PIL
Visual Studio 2015

Build

Compile deformable 3D convolution:

Cd to code/dcn
For Windows users, run cmd make.bat For Linux user, run bash make.sh The scripts will build D3D automatically and create some folders.
See test.py for example usage.

Datasets

Training dataset

Download the Vimeo dataset and put the images in code/data/Vimeo.
Cd to code/data/Vimeo and run generate_LR_Vimeo90K.m to generate training data as below:

  Vimeo
    └── sequences
           ├── 00001
           ├── 00002
           ├── ...
    └── LR_x4
           ├── 00001
           ├── 00002
           ├── ...		
    ├── readme.txt 
    ├── sep_trainlist.txt
    ├── sep_testlist.txt
    └── generate_LR_Vimeo90K.m

Test dataset

Download the dataset Vid4 and SPMC-11 dataset in https://pan.baidu.com/s/1PKZeTo8HVklHU5Pe26qUtw (Code: 4l5r) and put the folder in code/data.
(optional) You can also download Vid4 and SPMC-11 or other video datasets and prepare test data in code/data as below:

 data
  └── dataset_1
         └── scene_1
               └── hr    
                  ├── hr_01.png  
                  ├── hr_02.png  
                  ├── ...
                  └── hr_M.png    
               └── lr_x4
                  ├── lr_01.png  
                  ├── lr_02.png  
                  ├── ...
                  └── lr_M.png   
         ├── ...		  
         └── scene_M
  ├── ...    
  └── dataset_N

Results

Quantitative Results

Table 1. PSNR/SSIM achieved by different methods.

Table 2. Temporal consistency and computational efficiency achieved by different methods.

We have organized the Matlab code framework of Video Quality Assessment metric T-MOVIE and MOVIE. [Code]
Welcome to have a look and use our code.

Qualitative Results

Qualitative results achieved by different methods. Blue boxes represent the temporal profiles among different frames.

A demo video is available at https://wyqdatabase.s3-us-west-1.amazonaws.com/D3Dnet.mp4

Citiation

@article{D3Dnet,
  author = {Ying, Xinyi and Wang, Longguang and Wang, Yingqian and Sheng, Weidong and An, Wei and Guo, Yulan},
  title = {Deformable 3D Convolution for Video Super-Resolution},
  journal = {arXiv preprint arXiv:2004.02803},
  year = {2020},
}

Contact

Please contact us at yingxinyi18@nudt.edu.cn for any question.

shijun18/D3Dnet