Efficient Video Deblurring Guided by Motion Magnitude

Official repository of MMP-RNN (ECCV 2022) Paper
Estimate pixel-wise blur level (MMP) first and video deblur.

Requirements

Pytorch 1.8, Cuda 11.1
lmdb, tqdm, thop, scipy, opencv, scikit-image, tensorboard

Preparing ground truth motion magnitude prior

The ground truth MMP is generated from high-frequency sharp frames during exposure time. Optical flows between sharp frames are estimated by RAFT.

Learning MMP

Using a UNet-like structure to learn MMP.

MMP-RNN

Ultilizing MMP for video deblurring by merging into an RNN.

Citation

@inproceedings{wang2022MMP,
  title={Efficient Video Deblurring Guided by Motion Magnitude},
  author={Wang, Yusheng and Lu, Yunfan and Gao, Ye and Wang, Lin and Zhong, Zhihang and Zheng, Yinqiang and Yamashita, Atsushi},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2022}
}

Contact

Email of Yusheng Wang: wang@robot.t.u-tokyo.ac.jp