UEVD(ECCV 2022, Oral presentation)

This repository is for the ECCV 2022 paper, "Event-guided Deblurring of Unknown Exposure Time Videos".

[ArXiv] [ECCV2022] [Supp] [Oral(YouTube)] [Project]

Demo videos on real world blurry videos

Demo videos on the real-world event datset(Color-DVS)

Color-DVS

The first public event-based deblurring dataset. Our dataset contain diverse scene including real-world event using color-DAVIS camera.

You can download the raw-data(collected frame and events) from this google drive link

Also, you can download the processed data for handling unknown exposure time videos link

Installation

This code was tested with:

• pytorch 1.2.0
• CUDA 10.2
• Python 3.7
• Ubuntu 18.04 using TITAN RTX GPU

pip install -r requirements.txt
bash install.sh

Pre-trained model

To be uploaded

Test

python test_deblur_dvs.py --dataset 'dvs'

Training

python train_deblur_dvs.py --dataset 'dvs' --epochs 21 --batch_size 2 \
--test_batch_size 1 --use_multigpu True

Results

• The quantitative comparisons are attached as belows for a reference.

• The visual results of temporal activation map of the ETES modules on the vaious datasets.

Reference

Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon" Event-guided Deblurring of Unknown Exposure Time Videos", In ECCV, 2022.

BibTeX

@inproceedings{kim2022event,
  title={Event-guided Deblurring of Unknown Exposure Time Videos},
  author={Kim, Taewoo and Lee, Jeongmin and Wang, Lin and Yoon, Kuk-Jin},
  booktitle={European Conference on Computer Vision},
  pages={519--538},
  year={2022},
  organization={Springer}
}

Contact

If you have any question, please send an email me(intelpro@kaist.ac.kr)

License

The project codes and datasets can be used for research and education only.