Close Imitation of Expert Retouching for Black-and-White Photography (CVPR 2024 ACCEPTED !!)
Seunghyun Shin, Jisu Shin, Jihwan Bae, Inwook Shim and Hae-Gon Jeon
DeColorful-Net. We propose a DeColorfulNet, which is based on a DML framework with a hierarchical proxy-based loss and hierarchical bilateral grid network to mimic the experts’ retouching scheme
- Python >= 3.6
- PyTorch >= 1.0
- NVIDIA GPU + CUDA cuDNN
We propose the first aesthetic decolorization dataset which contains three different set retouched by experts.
Please download the dataset at https://drive.google.com/drive/folders/1pmvoaybrNvkAYeXa8bsx87t1Lr3Mfcb7?usp=drive_link .
Then, you should change "dataset_path" in json file where you save the dataset.
If you want to train DeColorful-Net Phase1, then download a subset of our dataset for train phase1 from https://drive.google.com/drive/folders/114AW6yDtFj8-6PSu8RwbLi03WImXYxu5?usp=drive_link in the same folder where you save the BW Adobe5k.
Our DeColorful-Net consists of two steps.
- Proxy-Generation
- Decolorization
Before train your model you should change some options with your settings which are listed in the form of json file.\
You can see json file with below command
cd ./workspace/Expert/Trial1
- Install python requirements:
pip install -r requirements.txt
python train_step1.py --ws Expert --exp Multi_Encoder --args json
ws: workspace
exp: experience space
args: which format to use options
Make sure to change directory of pretrained model from training phase 1.
python train_step2.py --ws Expert --exp Multi_Encoder --args json
Make sure to change directory of pretrained model from training phase 1 & 2.
python test.py --ws Expert --exp Multi_Encoder --args json
Our pretrained weights are released at: https://drive.google.com/drive/folders/192xs_tPeJmer-bnTFB1xng857xVfpxLn?usp=drive_link Make sure all weights should be on the experience folder ex) workspace/Expert/Trial1/step1.pth.tar
If you have any question, please feel free to contact us via seunghyuns98@gm.gist.ac.kr