Awesome
Unsupervised Exposure Correction (UEC) Documentation
Prerequisite
Dataset
To access the Afifi's Exposure-errors Dataset, please follow the instructions at this link.
For our Radiometry Correction Dataset, please check this link
Package Installation
Install the necessary Python packages by running the following command:
pip install -r requirements.txt
Pretrained Models
In the "checkpoints" directory, we have made available two ".pth" files, representing the outcomes of training on both Afifi's Exposure-errors Dataset and our Radiometry Correction Dataset, respectively.
Training
Make sure to replace dataset_root with the actual path where you downloaded the dataset.
python train.py --name exposure --model uec --dataset_mode exposure --load_size 448 --preprocess resize_and_crop --gpu_ids 2 --save_epoch_freq 1 --lr 1e-4 --beta1 0.9 --lr_policy step --lr_decay_iters 6574200 --dataset_root ../data/exposure_dataset/INPUT_IMAGES/
If you are using the Radiometry Correction Dataset, set --dataset_mode to fivek.
We removed TVLoss because we found the performance to be better without it. The PSNR result is as following:
| EV | -2 | -1 | 0 | +1 | +2 | +3 |
|---|---|---|---|---|---|---|
| w/ TVLoss | 22.577 | 20.528 | 18.336 | 17.820 | 15.752 | 15.138 |
| w/o TVLoss | 25.343 | 23.637 | 20.552 | 18.391 | 15.327 | 13.175 |
To reproduce the results from the paper, please run:
git reset --hard 9578ef19c250b349d2a247913af8e5e902e7f707
Testing
Replace dataset_root with the actual path for testing:
python test.py --name exposure-errors --model uec --dataset_mode fivektest --preprocess resize --load_size 256 --gpu_ids 2 --dataset_root ../data//exposure_dataset/test/ --ref_image_paths ../data/exposure_dataset/GT_IMAGES/a0001-jmac_DSC1459.jpg
The option --ref_image_paths is for choosing one reference for the final calibration.
Evaluation
We perform evaluation using PyIQA.
python inference_iqa_filelist.py -f ./results/file.txt -i ./results/exposure/test_latest/images/ -r <path-to-your-reference-images>
file.txt contains the mapping between input images and ground truth, , separated by tabs. If you need to resize the images, run the following command:
python script_name.py --img_size 256 --input_folder ./input_images --output_folder ./output_images
Acknowledgement
We brrow some code from BargainNet, NeurOP and Pix2PixHD. We would like to thanks the authors for sharing their excellent works.
License Agreement
- This repository with the provided code and pretrained model is available for non-commercial research purposes only.