EGVSR-PyTorch

VSR x4: EGVSR; Upscale x4: Bicubic Interpolation

EGVSR-PyTorch

Introduction

This is a PyTorch implementation of EGVSR: Efficcient & Generic Video Super-Resolution (VSR), using subpixel convolution to optimize the inference speed of TecoGAN VSR model. Please refer to the official implementation ESPCN and TecoGAN for more information.

Features

Unified Framework: This repo provides a unified framework for various state-of-the-art DL-based VSR methods, such as VESPCN, SOFVSR, FRVSR, TecoGAN and our EGVSR.
Multiple Test Datasets: This repo offers three types of video datasets for testing, i.e., standard test dataset -- Vid4, Tos3 used in TecoGAN and our new dataset -- Gvt72 (selected from Vimeo site and including more scenes).
Better Performance: This repo provides model with faster inferencing speed and better overall performance than prior methods. See more details in Benchmarks section.

Dependencies

Ubuntu >= 16.04
NVIDIA GPU + CUDA & CUDNN
Python 3
PyTorch >= 1.0.0
Python packages: numpy, matplotlib, opencv-python, pyyaml, lmdb (requirements.txt & req.txt)
(Optional) Matlab >= R2016b

Datasets

A. Training Dataset

Download the official training dataset based on the instructions in TecoGAN-TensorFlow, rename to VimeoTecoGAN and then place under ./data.

B. Testing Datasets

Vid4 -- Four video sequences: city, calendar, foliage and walk;
Tos3 -- Three video sequences: bridge, face and room;
Gvt72 -- Generic VSR Test Dataset: 72 video sequences (including natural scenery, culture scenery, streetscape scene, life record, sports photography, etc, as shown below)

You can get them at ⏬ 百度网盘 (提取码:8tqc) and put them into 📁 Datasets. The following shows the structure of the above three datasets.

data
  ├─ Vid4
    ├─ GT                # Ground-Truth (GT) video sequences
      └─ calendar
        ├─ 0001.png
        └─ ...
    ├─ Gaussian4xLR      # Low Resolution (LR) video sequences in gaussian degradation and x4 down-sampling
      └─ calendar
        ├─ 0001.png
        └─ ...
  └─ ToS3
    ├─ GT
    └─ Gaussian4xLR
  └─ Gvt72
    ├─ GT
    └─ Gaussian4xLR

Benchmarks

Experimental Environment

	Version	Info.
System	Ubuntu 18.04.5 LTS	X86_64
CPU	Intel i9-9900	3.10GHz
GPU	Nvidia RTX 2080Ti	11GB GDDR6
Memory	DDR4 2666	32GB×2

A. Test on Vid4 Dataset

1.LR 2.VESPCN 3.SOFVSR 4.DUF 5.Ours:EGVSR 6.GT
Objective metrics for visual quality evaluation[1]

B. Test on Tos3 Dataset

1.VESPCN 2.SOFVSR 3. FRVSR 4.TecoGAN 5.Ours:EGVSR 6.GT

C. Test on Gvt72 Dataset

1.LR 2.VESPCN 3.SOFVSR 4.DUF 5.Ours:EGVSR 6.GT
Objective metrics for visual quality and temporal coherence evaluation[1]

D. Optical-Flow based Motion Compensation

Please refer to FLOW_walk, FLOW_foliage and FLOW_city.

E. Comprehensive Performance

Comparison of various SOTA VSR model on video quality score and speed performance[3]

^[1] ⬇️:smaller value for better performance, ⬆️: on the contrary; Red: stands for Top1, Blue: Top2. ^[2] The calculation formula of video quality score considering both spatial and temporal domain, using lambda1=lambda2=lambda3=1/3. ^[3] FLOPs & speed are computed on RGB with resolution 960x540 to 3840x2160 (4K) on NVIDIA GeForce GTX 2080Ti GPU.

License & Citations

This EGVSR project is released under the MIT license. See more details in LICENSE. The provided implementation is strictly for academic purposes only. If EGVSR helps your research or work, please consider citing EGVSR. The following is a BibTeX reference:

@misc{thmen2021egvsr,
  author =       {Yanpeng Cao, Chengcheng Wang, Changjun Song, Yongming Tang and He Li},
  title =        {EGVSR},
  howpublished = {\url{https://github.com/Thmen/EGVSR}},
  year =         {2021}
}

Yanpeng Cao, Chengcheng Wang, Changjun Song, Yongming Tang and He Li. EGVSR. https://github.com/Thmen/EGVSR, 2021.

Acknowledgements

This code is built on the following projects. We thank the authors for sharing their codes.

Thmen/EGVSR