RSCD (BS-RSCD & JCD)
Towards Rolling Shutter Correction and Deblurring in Dynamic Scenes (CVPR2021)
by Zhihang Zhong, Yinqiang Zheng, Imari Sato
We contributed the first real-world dataset (BS-RSCD) and end-to-end model (JCD) for joint rolling shutter correction and deblurring tasks.
We collected the data samples using the proposed beam-splitter acquisition system as below:
In the near future, we will add more data samples with larger distortion to the dataset ...
If you are interested in real-world datasets for pure deblurring tasks, please refer to ESTRNN & BSD.
Prerequisites
Install the dependent packages:
conda create -n rscd python=3.8
conda activate rscd
sh install.sh
Download lmdb files of BS-RSCD
(or Fastec-RS for RSC tasks).
(PS, for how to create lmdb file, you can refer to ./data/create_rscd_lmdb.ipynb)
Training
Please specify the <path> (e.g. "./dataset/ ") where you put the dataset file or change the default value in "
./para/paramter.py".
Train JCD on BS-RSCD:
python main.py --data_root <path> --model JCD --dataset rscd_lmdb --video
Train JCD on Fastec-RS:
python main.py --data_root <path> --model JCD --dataset fastec_rs_lmdb --video
Testing
Please download checkpoints and
unzip it under the main directory.
Run the pre-trained model on BS-RSCD:
python main.py --test_only --dataset rscd_lmdb --test_checkpoint ./checkpoints/JCD_BS-RSCD.tar --video
Inference for video file:
python video_inference.py --src <input_path> --dst <output_path> --checkpoint ./checkpoints/JCD_BS-RSCD.tar
Citing
If BS-RSCD and JCD are useful for your research, please consider citing:
@InProceedings{Zhong_2021_Towards,
title={Towards Rolling Shutter Correction and Deblurring in Dynamic Scenes},
author={Zhong, Zhihang and Zheng, Yinqiang and Sato, Imari},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year={2021}
}