Awesome

Neural Color Operators for Sequential Image Retouching (ECCV2022)

Yili Wang, Xin Li, Kun Xu, Dongliang He, Qi Zhang, Fu Li, Errui Ding

[arXiv] [project] [doi]

[Paddle Implementation](Offical)

[Pytorch Implementation]

[Jittor Implementation]

<img src="figures/advantage.png"> Left: Compared with previous state-of-the-art methods, NeurOp achieves superior performance with only 28k parameters (~75% of CSRNet). Right: Strength Controllability Results. Our method can directly change the retouching output with intuitive control (i.e. directly modify the scalar strengths) <img src="figures/result.png">

Datasets

Pretrain data to initialize our neurOps is hosted on 百度网盘 (code:pld9).

MIT-Adobe FiveK & PPR10K

We host all these data in 百度网盘 (code:jvvq)

There are two preprocessed versions of MIT-Adobe FiveK, in our paper, we refer them as MIT-Adobe FiveK-Dark (originally provided by CSRNet) and MIT-Adobe FiveK-Lite (originally provided by Distort-and-Recover).
The official PPR10K dataset link is here.

Get Started

Clone this repo

git clone https://github.com/amberwangyili/neurop

Download the Dataset from 百度网盘 (code:jvvq) and unzip in project folder

tree -L 2 neurop/datasets
# the output should be like the following:
datasets/
├── dataset-dark
│   ├── testA
│   ├── testB
│   ├── trainA
│   └── trainB
├── dataset-init
│   ├── BC
│   ├── EX
│   └── VB
├── dataset-lite
│   ├── testA
│   ├── testB
│   ├── trainA
│   └── trainB
└── dataset-ppr
    ├── ppr-a
    ├── ppr-b
    ├── ppr-c
    ├── testA
    ├── testM
    ├── trainA
    └── trainM

Install Dependencies

cd neurop
pip install -r requirements.txt

Test

We provide pretrained model weights for MIT-Adobe FiveK and PPR10K in pretrain_models

Run command:

python test.py -config ./configs/test/<configuaration-name>.yaml

The evaluation results will be in the neurop/results folder

Train

Initialization individual neural color operators:

python train.py -config ./configs/init_neurop.yaml

Finetune with strength predictors:

python train.py -config ./configs/train/<configuration-name>.yaml

BibTex

If you find neurOp useful in your research, please use the following BibTeX entry.

    @inproceedings{wang2022neurop,
    author = {Wang, Yili and Li, Xin and Xu, Kun and He, Dongliang and Zhang, Qi and Li, Fu and Ding, Errui},
    title = {Neural Color Operators for Sequential Image Retouching},
    year = {2022},
    isbn = {978-3-031-19800-7},
    publisher = {Springer-Cham},
    url = {https://doi.org/10.1007/978-3-031-19800-7_3},
    doi = {10.1007/978-3-031-19800-7_3},
    booktitle = {Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XIX},
    numpages = {14},
    }

Acknowledgement

NeurOp is licensed under a MIT License.