Skip to content
/ TCA Public

Repository for paper Temporal consistency learning for video super-resolution.

License

Notifications You must be signed in to change notification settings

mrluin/TCA

Repository files navigation

Learning Temporal Consistency for Video Super-Resolution

official implementation

GIF Visualization of BasicVSR and Ours

GIF Visualization of animated figure, Fig.1 and Fig.7-Fig9, in our paper.

Fig. 1. Visual comparison of BasicVSR (top-right) and our TCA learning (bottom-right). The left if HR video. BasicVSR suffers from flickering artifacts, while our TCA learning can better keep temporal consistency of video results.

Fig. 7. Visual comparison of BasicVSR (left, red rectangle) and our TCA learning (right, cyan-blue rectangle). In rectangle region, BasicVSR suffers from flickering artifacts, while in our TCA learning flickering artifacts are effectively alleviated.

Fig. 8. Visual comparison of BasicVSR (left, top red rectangle) and our TCA learning (right, bottom cyan-blue rectangle). In rectangle region, BasicVSR suffers from flickering artifacts, while in our TCA learning flickering artifacts are effectively alleviated.

Fig. 9. Visual comparison of BasicVSR (left, top red rectangle) and our TCA learning (right, bottom cyan-blue rectangle). In rectangle region, BasicVSR suffers from flickering artifacts, while in our TCA learning flickering artifacts are effectively alleviated.

Dataset

Test data includes REDS4, Vid4, SPMC30, Vimeo90K-T and UDM10. Before evaluation, please download corresponding testing and put them in given paths.

Type Download URL Path
REDS4 REDS4 download ./dataset/REDS4
SPMC30 SPMC30 download ./dataset/SPMC30
UDM10 UDM10 download ./dataset/UDM10
Vid4 Vid4 download ./dataset/Vid4
Vimeo90K-T Vimeo90K-T download ./dataset/Vimeo90K-T

For BI degradation, we adopt matlab imresize function.

For BD degradation, we adopt BD_degradation.m.

Note that LR of SPMC30 has pixel shift, you should use standard imresize function to obtain LR frames.

Code

installation

# create conda env
conda create -n tca python=3.7

# install pytorch, torchvision, cupy
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=9.2 -c pytorch
conda install -c conda-forge cupy cudatoolkit=10.2

# install other dependencies
pip install pyyaml, wandb, opencv-python, tqdm, lpips

Checkpoints

Download checkpoints from Google Drive, and put them in ./chekpoints.

Evaluation

Inference model on testing data:

python inference.py --opt ./options/tca_reds_bi4x_reds4_inference.yml

python inference.py --opt ./options/tca_vimeo90k_bi4x_vid4_vimeo90kt_spmc30_inference.yml

python inference.py --opt ./options/tca_vimeo90k_bd4x_udm10_vid4_vimeo90kt_spmc30.yml

Metrics

Calculate PSNR, SSIM, LPIPS and flow warping error:

python metrics_calculator.py

python flow_warping_error_calculator.py

Qualitative Comparison

Animated comparison results please refer to Google Drive.

License

This project is released under the Apache 2.0 license.

Acknowledgement

This implementation largely depends on BasicSR. Thanks for the excellent codebase!

About

Repository for paper Temporal consistency learning for video super-resolution.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published