KAIST-VICLab / 3DSRnet

Official repository of 3DSRnet (ICIP 2019)

3DSRnet

Official repository of 3DSRnet (ICIP2019)

We provide the training and test code along with the trained weights and the dataset (train+test) used for 3DSRnet. The result videos on the Vidset4 benchmark are provided here. (Results of Bicubic, VSRnet, VESPCN, 3DSRnet and GT provided)

If you find this repository useful, please consider citing our paper.

Reference:

Soo Ye Kim, Jeongyeon Lim, Taeyoung Na, Munchurl Kim. Video Super-Resolution Based on 3D-CNNS with Consideration of Scene Change. IEEE International Conference on Image Processing, 2019. Electronic Poster

Extended paper on arXiv:

Soo Ye Kim, Jeongyeon Lim, Taeyoung Na, Munchurl Kim. 3DSRnet: Video Super-resolution using 3D Convolutional Neural Networks. arXiv: 1812.09079

Requirements

Our code is implemented using MatConvNet. (MATLAB required)

Appropriate installations of MatConvNet is necessary via the official website.
Detailed instructions on installing MatConvNet can be found here.

The 3D convolution layer is implemented based on pengsun's mex implementation in GitHub.
MexConv3D must be installed prior to executing any of the provided source code.

The code was tested under the following setting:

• MATLAB 2017a
• MatConvNet 1.0-beta25
• CUDA 9.0, 10.0
• cuDNN 7.1.4
• NVIDIA TITAN Xp GPU

Test code

Quick Start (Video SR Benchmark)

1. Download the source code in a directory of your choice <source_path>.
2. Download the test dataset (Vid4) from this link and place the 'test' folder in <source_path>/data
3. Place the files in <source_path>/+dagnn/ to <MatConvNet>/matlab/+dagnn
4. Run test.m

Description

We provide the pre-trained weights for the x2, x3 and x4 models in <source_path>/net.
The test dataset (Vid4) can be downloaded from here. With test.m, the pre-trained models can be evaluated on the Vid4 benchmark.

Remarks

• You can change the SR scale factor (2, 3 or 4) by modifying the 'scale' parameter in the initial settings.
• You can change the video sequence by modifying the 'sequence_name' parameter in the initial settings.
• When you run this code, evaluation will be performed on PSNR and the .png prediction files will be saved in <source_path>/pred/

Quick Start (with SF subnet)

1. Download the source code in a directory of your choice <source_path>.
2. Place the files in <source_path>/+dagnn/ to <MatConvNet>/matlab/+dagnn
3. Run test_SF_subnet.m or test_SF_SR.m

Description

• The pre-trained weights of the SF subnet is given in <source_path>/net.
• Four samples of data containing a scene boundary after frame 1, 2, 3 and 4, and a sample containing no scene change are provided in <source_path>/data/SF_subnet.
• With test_SF_subnet.m, you can test the scene boundary detection of the SF subnet for the given sample data.
• In test_SF_SR.m, the whole pipeline of detecting the scene boundary, replacing the different scene frames, and finally inferring the video SR network is implemented. When you run this code, the .png prediction files will be saved in <source_path>/pred/SF_SR. You can change the SR scale factor (2, 3 or 4) by modifying the 'scale' parameter in the initial settings.

Training code

Quick Start

1. Download the source code in a directory of your choice <source_path>.
2. Download the train dataset from here and place the 'train' folder in <source_path>/data
3. Place the files in <source_path>/+dagnn/ to <MatConvNet>/matlab/+dagnn
4. Run train.m

Description

This code (train.m) trains the video SR subnet. The 3D-CNN model of the video SR subnet is specified in net.m. The train dataset can be downloaded from here.

Remarks

• You can change the SR scale factor (2, 3 or 4) by modifying the 'scale' parameter.
• The trained weights will be saved in <source_path>/net/net_xscale

Contact

Please contact me via email ([e-mail]) for any problems regarding the released code.
Note: We plan to provide the source code for the scene change module (SF subnet) in the near future.