Conditional Sequential Modulation for Efficient Global Image Retouching Paper Link
By Jingwen He*, Yihao Liu*, Yu Qiao, and Chao Dong (* indicates equal contribution)
Left: Compared with existing state-of-the-art methods, our method achieves superior performance with extremely few parameters (1/13 of HDRNet and 1/250 of White-Box). The diameter of the circle represents the amount of trainable parameters. Right: Image retouching examples. The first row shows smooth transition effects between different styles (expert A to B) by image interpolation. In the second row, we use image interpolation to control the retouching strength from input image to the automatic retouched result. We denote the interpolation coefficient α for each image.@article{he2020conditional,
title={Conditional Sequential Modulation for Efficient Global Image Retouching},
author={He, Jingwen and Liu, Yihao and Qiao, Yu and Dong, Chao},
journal={arXiv preprint arXiv:2009.10390},
year={2020}
}
- Python 3 (Recommend to use Anaconda)
- PyTorch >= 1.0
- NVIDIA GPU + CUDA
- Python packages:
pip install numpy opencv-python lmdb pyyaml
- TensorBoard:
- PyTorch >= 1.1:
pip install tb-nightly future
- PyTorch == 1.0:
pip install tensorboardX
- PyTorch >= 1.1:
Here, we provide the preprocessed datasets: MIT-Adobe FiveK dataset, which contains both training pairs and testing pairs.
- training pairs: {GT: expert_C_train; Input: raw_input_train}
- testing pairs: {GT: expert_C_test; Input: raw_input_test}
- Modify the configuration file
options/test/test_Enhance.yml
. e.g.,dataroot_GT
,dataroot_LQ
, andpretrain_model_G
. (We provide a pretrained model inexperiments/pretrain_models/csrnet.pth
) - Run command:
python test_CSRNet.py -opt options/test/test_Enhance.yml
- Modify the python file
calculate_metrics.py
:input_path
,GT_path
(Line 139, 140). Then run:
python calculate_metrics.py
- Modify the configuration file
options/train/train_Enhance.yml
. e.g.,dataroot_GT
,dataroot_LQ
. - Run command:
python train.py -opt options/train/train_Enhance.yml
- This code is based on mmsr.
- Thanks Yihao Liu for part of this work.