Skip to content
/ HDRUNet Public

CVPR2021 Workshop - HDRUNet: Single Image HDR Reconstruction with Denoising and Dequantization.

License

Notifications You must be signed in to change notification settings

chxy95/HDRUNet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

34 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HDRUNet [Paper Link]

HDRUNet: Single Image HDR Reconstruction with Denoising and Dequantization

Xiangyu Chen, Yihao Liu, Zhengwen Zhang, Yu Qiao and Chao Dong

We won the second place in NTIRE2021 HDR Challenge (Track1: Single Frame). The paper is accepted to CVPR2021 Workshop.

BibTeX

@InProceedings{chen2021hdrunet,
    author    = {Chen, Xiangyu and Liu, Yihao and Zhang, Zhengwen and Qiao, Yu and Dong, Chao},
    title     = {HDRUNet: Single Image HDR Reconstruction With Denoising and Dequantization},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {354-363}
}

Overview

Overview of the network:

Overview of the loss function:

Tanh_L1(Y, H) = |Tanh(Y) - Tanh(H)|

Getting Started

  1. Dataset
  2. Configuration
  3. How to test
  4. How to train
  5. Visualization

Dataset

Register a codalab account and log in, then find the download link on this page:

https://competitions.codalab.org/competitions/28161#participate-get-data

It is strongly recommended to use the data provided by the competition organizer for training and testing, or you need at least a basic understanding of the competition data. Otherwise, you may not get the desired result.

Configuration

pip install -r requirements.txt

How to test

  • Modify dataroot_LQ and pretrain_model_G (you can also use the pretrained model which is provided in the ./pretrained_model) in ./codes/options/test/test_HDRUNet.yml, then run
cd codes
python test.py -opt options/test/test_HDRUNet.yml

The test results will be saved to ./results/testset_name.

How to train

  • Prepare the data. Modify input_folder and save_folder in ./scripts/extract_subimgs_single.py, then run
cd scripts
python extract_subimgs_single.py
  • Modify dataroot_LQ and dataroot_GT in ./codes/options/train/train_HDRUNet.yml, then run
cd codes
python train.py -opt options/train/train_HDRUNet.yml

The models and training states will be saved to ./experiments/name.

Visualization

In ./scripts, several scripts are available. data_io.py and metrics.py are provided by the competition organizer for reading/writing data and evaluation. Based on these codes, I provide a script for visualization by using the tone-mapping provided in metrics.py. Modify paths of the data in ./scripts/tonemapped_visualization.py and run

cd scripts
python tonemapped_visualization.py

to visualize the images.

Acknowledgment

The code is inspired by BasicSR.

About

CVPR2021 Workshop - HDRUNet: Single Image HDR Reconstruction with Denoising and Dequantization.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages