Awesome

Recurrent Video Restoration Transformer with Guided Deformable Attention (RVRT, NeurlPS2022)

arxiv | supplementary | pretrained models | visual results

<a href="https://colab.research.google.com/gist/JingyunLiang/23502e2c65d82144219fa3e3322e4fc3/rvrt-demo-on-video-restoration.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google colab logo"></a>

This repository is the official PyTorch implementation of "Recurrent Video Restoration Transformer with Guided Deformable Attention" (arxiv, supp, pretrained models, visual results). RVRT achieves state-of-the-art performance with balanced model size, testing memory and runtime in

• video SR (REDS, Vimeo90K, Vid4, UDM10)
• video deblurring (GoPro, DVD)
• video denoising (DAVIS, Set8)

:rocket: :rocket: :rocket: News:

• June. 8, 2022: See more related works in vision restoration as follows:

Topic	Title
transformer-based image/video restoration:fire:	SwinIR: Image Restoration Using Swin Transformer <br /> VRT: A Video Restoration Transformer
real-world image SR/denoising	Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis <br />Designing a Practical Degradation Model for Deep Blind Image Super-Resolution, ICCV2021
blind image SR	Flow-based Kernel Prior with Application to Blind Super-Resolution, CVPR2021 <br /> Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution, ICCV2021
generative image SR and image rescaling	Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling, ICCV2021

Video restoration aims at restoring multiple high-quality frames from multiple low-quality frames. Existing video restoration methods generally fall into two extreme cases, i.e., they either restore all frames in parallel or restore the video frame by frame in a recurrent way, which would result in different merits and drawbacks. Typically, the former has the advantage of temporal information fusion. However, it suffers from large model size and intensive memory consumption; the latter has a relatively small model size as it shares parameters across frames; however, it lacks long-range dependency modeling ability and parallelizability. In this paper, we attempt to integrate the advantages of the two cases by proposing a recurrent video restoration transformer, namely RVRT. RVRT processes local neighboring frames in parallel within a globally recurrent framework which can achieve a good trade-off between model size, effectiveness, and efficiency. Specifically, RVRT divides the video into multiple clips and uses the previously inferred clip feature to estimate the subsequent clip feature. Within each clip, different frame features are jointly updated with implicit feature aggregation. Across different clips, the guided deformable attention is designed for clip-to-clip alignment, which predicts multiple relevant locations from the whole inferred clip and aggregates their features by the attention mechanism. Extensive experiments on video super-resolution, deblurring, and denoising show that the proposed RVRT achieves state-of-the-art performance on benchmark datasets with balanced model size, testing memory and runtime.

Contents

1. Requirements
2. Quick Testing
3. Training
4. Results
5. Citation
6. License and Acknowledgement

Requirements

• Python 3.8, PyTorch >= 1.9.1
• Requirements: see requirements.txt
• Platforms: Ubuntu 18.04, cuda-11.1

Quick Testing

Following commands will download pretrained models and test datasets automatically (except Vimeo-90K testing set). If out-of-memory, try to reduce --tile at the expense of slightly decreased performance.

You can also try to test it on Colab <a href="https://colab.research.google.com/gist/JingyunLiang/23502e2c65d82144219fa3e3322e4fc3/rvrt-demo-on-video-restoration.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google colab logo"></a>, but the results may be slightly different due to --tile difference.

# download code
git clone https://github.com/JingyunLiang/RVRT
cd RVRT
pip install -r requirements.txt

# 001, video sr trained on REDS, tested on REDS4
python main_test_rvrt.py --task 001_RVRT_videosr_bi_REDS_30frames --folder_lq testsets/REDS4/sharp_bicubic --folder_gt testsets/REDS4/GT --tile 100 128 128 --tile_overlap 2 20 20

# 002, video sr trained on Vimeo (bicubic), tested on Vid4 and Vimeo
python main_test_rvrt.py --task 002_RVRT_videosr_bi_Vimeo_14frames --folder_lq testsets/Vid4/BIx4 --folder_gt testsets/Vid4/GT --tile 0 0 0 --tile_overlap 2 20 20
python main_test_rvrt.py --task 002_RVRT_videosr_bi_Vimeo_14frames --folder_lq testsets/vimeo90k/vimeo_septuplet_matlabLRx4/sequences --folder_gt testsets/vimeo90k/vimeo_septuplet/sequences --tile 0 0 0 --tile_overlap 0 20 20

# 003, video sr trained on Vimeo (blur-downsampling), tested on Vid4, UDM10 and Vimeo
python main_test_rvrt.py --task 003_RVRT_videosr_bd_Vimeo_14frames --folder_lq testsets/Vid4/BDx4 --folder_gt testsets/Vid4/GT --tile 0 0 0 --tile_overlap 2 20 20
python main_test_rvrt.py --task 003_RVRT_videosr_bd_Vimeo_14frames --folder_lq testsets/UDM10/BDx4 --folder_gt testsets/UDM10/GT --tile 0 0 0 --tile_overlap 2 20 20
python main_test_rvrt.py --task 003_RVRT_videosr_bd_Vimeo_14frames --folder_lq testsets/vimeo90k/vimeo_septuplet_BDLRx4/sequences --folder_gt testsets/vimeo90k/vimeo_septuplet/sequences --tile 0 0 0 --tile_overlap 0 20 20

# 004, video deblurring trained and tested on DVD
python main_test_rvrt.py --task 004_RVRT_videodeblurring_DVD_16frames --folder_lq testsets/DVD10/test_GT_blurred --folder_gt testsets/DVD10/test_GT --tile 0 256 256 --tile_overlap 2 20 20

# 005, video deblurring trained and tested on GoPro
python main_test_rvrt.py --task 005_RVRT_videodeblurring_GoPro_16frames --folder_lq testsets/GoPro11/test_GT_blurred --folder_gt testsets/GoPro11/test_GT --tile 0 256 256 --tile_overlap 2 20 20

# 006, video denoising trained on DAVIS (noise level 0-50) and tested on Set8 and DAVIS
python main_test_rvrt.py --task 006_RVRT_videodenoising_DAVIS_16frames --sigma 50 --folder_lq testsets/Set8 --folder_gt testsets/Set8 --tile 0 256 256 --tile_overlap 2 20 20
python main_test_rvrt.py --task 006_RVRT_videodenoising_DAVIS_16frames --sigma 50  --folder_lq testsets/DAVIS-test --folder_gt testsets/DAVIS-test --tile 0 256 256 --tile_overlap 2 20 20

# test on your own datasets (an example)
python main_test_rvrt.py --task 001_RVRT_videosr_bi_REDS_30frames --folder_lq testsets/your/own --tile 0 0 0 --tile_overlap 2 20 20

All visual results of RVRT can be downloaded here.

Training

The training and testing sets are as follows (see the supplementary for a detailed introduction of all datasets). For better I/O speed, use create_lmdb.py to convert .png datasets to .lmdb datasets.

Note: You do NOT need to prepare the datasets if you just want to test the model. main_test_rvrt.py will download the testing set automaticaly.

The training code will be put in KAIR.

Results

We achieved state-of-the-art performance on video SR, video deblurring and video denoising. Detailed results can be found in the paper.

Video Super-Resolution

Video Deblurring

Video Denoising

Citation

@article{liang2022rvrt,
    title={Recurrent Video Restoration Transformer with Guided Deformable Attention},
    author={Liang, Jingyun and Fan, Yuchen and Xiang, Xiaoyu and Ranjan, Rakesh and Ilg, Eddy  and Green, Simon and Cao, Jiezhang and Zhang, Kai and Timofte, Radu and Van Gool, Luc},
    journal={arXiv preprint arXiv:2206.02146},
    year={2022}
}

License and Acknowledgement

This project is released under the CC-BY-NC license. We refer to codes from KAIR, BasicSR, Video Swin Transformer and mmediting. Thanks for their awesome works. The majority of RVRT is licensed under CC-BY-NC, however portions of the project are available under separate license terms: KAIR is licensed under the MIT License, BasicSR, Video Swin Transformer and mmediting are licensed under the Apache 2.0 license.