Bringing Events into Video Deblurring with Non consecutively Blurry Frames （ICCV2021）

[PDF]

Introduction

Existing video deblurring methods assume consecutively blurry frames, while neglecting the fact that sharp frames usually appear nearby blurry frame. In this paper, we develop a principled framework D2Nets for video deblurring to exploit non-consecutively blurry frames, and propose a flexible event fusion module (EFM) to bridge the gap between event-driven and video deblurring.

Prerequisites

Python >= 3.6, PyTorch >= 1.1.0
Requirements: opencv-python, numpy, matplotlib, imageio, scikit-image, tqdm, cupy
Platforms: Ubuntu 20.04, cuda-10.0, 2*2080Ti

Updatas

[2022-01-14] Our D2Net model trained on consecutively blurry GOPRO dataset can be download Here[password: blrf]

Datasets

GOPRO_Random: To satisfy our assumption that sharp frames exist in a blurry video, we generate non-consecutively blurry frames in a video by randomly averaging adjacent sharp frames, i.e., the average number is randomly chosen from 1 to 15. And we assume that a generated frame Bi is sharp if the number of averaging frames is smaller than 5, i.e., label is 1, otherwise label is 0. It is worth noting that we randomly generate 50% blurry frames in a video, while the other 50% frames are sharp, without constraining that there must be 2 sharp ones in consecutive 7 frame.

Dataset Organization Form

|--dataset
    |--blur  
        |--video 1
            |--frame 1
            |--frame 2
                ：  
        |--video 2
            :
        |--video n
    |--gt
        |--video 1
            |--frame 1
            |--frame 2
                ：  
        |--video 2
         :
        |--video n
    |--Event
        |--video 1
            |--frame 1
            |--frame 2
                ：  
        |--video 2
         :
        |--video n
    |--label
        |--video 1
        |--video 2
         :
        |--video n

Download

Please download the testing datasets from BaiduYun[password: ho6f] and the training datasets from BaiduYun[password: ju3q]. And pretrained PWCFlow model can be downloaded Here[password: wkt0] and Our D2Net model trained on non-consecutively blurry GOPRO dataset can be download Here[password: 16fr]

(i) If you have downloaded the pretrained models，please put PWC_Flow model to './pretrain_models' and D2Net model to './code/logs', respectively.

(ii) If you have downloaded the datasets，please put them to './dataset'.

Getting Started

1) Testing

You need first choose w/ or w/o event data in '.sh'.

bash Inference.sh

The results on GOPRO_Random are also available at BaiduYun[password: d3as]

Note: The results in our paper is testing on 4X down sampling GOPRO due to the large event data. Now we have retrained our method on original resolution of GOPRO.

Metrics(PSNR/SSIM) calculating codes are Here. Average PSNR/SSIM values on GOPRO_Random dataset:

Dataset	DMPHN	STFAN	CDVD-TSP	ESTRNN	D2Net	D2Net*
GOPRO_Random	32.09/0.897	31.76/0.873	33.76/0.925	33.52/0.912	35.53/0.948	35.72/0.951

For testing consecutively blurry dataset, please put D2Net official model to './code/offical_logs'. You need first choose w/ or w/o event data in '.sh'.

bash Inference_official.sh

2) Training

Without event data:

python main_d2net.py --template D2NET

With event data:

python main_d2net_event.py --template D2NET_EVENT

Please check the path for you dataset.

Cite

If you use any part of our code, or D2Net and non consecutively blurry dataset are useful for your research, please consider citing:

  @InProceedings{Shang_2021_ICCV,
      author    = {Shang, Wei and Ren, Dongwei and Zou, Dongqing and Ren, Jimmy S. and Luo, Ping and Zuo, Wangmeng},
      title     = {Bringing Events Into Video Deblurring With Non-Consecutively Blurry Frames},
      booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
      month     = {October},
      year      = {2021},
      pages     = {4531-4540}
  }

Acknowledgements

This code is built on CDVD-TSP. We thank the authors for sharing the codes.

shangwei5/D2Net