Code for HDR Video Reconstruction
HDR Video Reconstruction: A Coarse-to-fine Network and A Real-world Benchmark Dataset (ICCV 2021)
Guanying Chen, Chaofeng Chen, Shi Guo, Zhetong Liang, Kwan-Yee K. Wong, Lei Zhang
Table of Contents
Overview:
We provide testing and training code. Details of the training and testing dataset can be found in DeepHDRVideo-Dataset. Datasets and the trained models can be download in Google Drive or BaiduYun (TODO).
Dependencies
This model is implemented in PyTorch and tested with Ubuntu (14.04 and 16.04) and Centos 7.
- Python 3.7
- PyTorch 1.10 and torchvision 0.30
You are highly recommended to use Anaconda and create a new environment to run this code. The following is an example procedure to install the dependencies.
# Create a new python3.7 environment named hdr
conda create -n hdr python=3.7
# Activate the created environment
source activate hdr
pip install -r requirements.txt
# Build deformable convolutional layer, tested with pytorch 1.1, g++5.5, and cuda 9.0
cd extensions/dcn/
python setup.py develop
# Please refer to https://github.com/xinntao/EDVR if you have difficulty in building this module
Testing
Please first go through DeepHDRVideo-Dataset to familiarize yourself with the testing dataset.
The trained models can be found in Google Drive (Models/). Download and place it to data/models/.
Testing on the synthetic test dataset
The synthetic test dataset can be found in Google Drive (/Synthetic_Dataset/HDR_Synthetic_Test_Dataset.tgz). Download and unzip it to data/. Note that we donot perform global motion alignment for this synthetic dataset.
# Test our method on two-exposure data. Results can be found in data/models/CoarseToFine_2Exp/
python run_model.py --gpu_ids 0 --model hdr2E_flow2s_model \
--benchmark syn_test_dataset --bm_dir data/HDR_Synthetic_Test_Dataset \
--mnet_name weight_net --mnet_checkp data/models/CoarseToFine_2Exp/weight_net.pth --fnet_checkp data/models/CoarseToFine_2Exp/flow_net.pth --mnet2_checkp data/models/CoarseToFine_2Exp/refine_net.pth
# Test our method on three-exposure data. The results can be found in data/models/CoarseToFine_3Exp/
python run_model.py --gpu_ids 0 --model hdr3E_flow2s_model \
--benchmark syn_test_dataset --bm_dir data/HDR_Synthetic_Test_Dataset \
--mnet_name weight_net --mnet_checkp data/models/CoarseToFine_3Exp/weight_net.pth --fnet_checkp data/models/CoarseToFine_3Exp/flow_net.pth --mnet2_checkp data/models/CoarseToFine_3Exp/refine_net.pth
Testing on the TOG13 dataset
Please download this dataset from TOG13_Dynamic_Dataset.tgz and unzip to data/. Normally when testing on a video, we have to first compute the similarity transformation matrices between neighboring frames using the following commands.
# However, this is optional as the downloaded dataset already contains the require transformation matrices for each scene in Affine_Trans_Matrices/.
python utils/compute_nbr_trans_for_video.py --in_dir data/TOG13_Dynamic_Dataset/ --crf data/TOG13_Dynamic_Dataset/BaslerCRF.mat --scene_list 2Exp_scenes.txt
python utils/compute_nbr_trans_for_video.py --in_dir data/TOG13_Dynamic_Dataset/ --crf data/TOG13_Dynamic_Dataset/BaslerCRF.mat --scene_list 3Exp_scenes.txt
# Test our method on two-exposure data. The results can be found in data/models/CoarseToFine_2Exp/
# Specify the testing scene with --test_scene. Available options are Ninja-2Exp-3Stop WavingHands-2Exp-3Stop Skateboarder2-3Exp-2Stop ThrowingTowel-2Exp-3Stop
python run_model.py --gpu_ids 0 --model hdr2E_flow2s_model \
--benchmark tog13_online_align_dataset --bm_dir data/TOG13_Dynamic_Dataset --test_scene ThrowingTowel-2Exp-3Stop --align \ --mnet_name weight_net --fnet_checkp data/models/CoarseToFine_2Exp/flow_net.pth --mnet_checkp data/models/CoarseToFine_2Exp/weight_net.pth --mnet2_checkp data/models/CoarseToFine_2Exp/refine_net.pth
# To test on a specific scene, you can use the --test_scene argument, e.g., "--test_scene ThrowingTowel-2Exp-3Stop".
# Test our method on three-exposure data. The results can be found in data/models/CoarseToFine_3Exp/
# Specify the testing scene with --test_scene. Available options are Cleaning-3Exp-2Stop Dog-3Exp-2Stop CheckingEmail-3Exp-2Stop Fire-2Exp-3Stop
python run_model.py --gpu_ids 0 --model hdr3E_flow2s_model \
--benchmark tog13_online_align_dataset --bm_dir data/TOG13_Dynamic_Dataset --test_scene Dog-3Exp-2Stop --align \
--mnet_name weight_net --fnet_checkp data/models/CoarseToFine_3Exp/flow_net.pth --mnet_checkp data/models/CoarseToFine_3Exp/weight_net.pth --mnet2_checkp data/models/CoarseToFine_3Exp/refine_net.pth
Testing on the captured static dataset
The global motion augmented static dataset can be found in Google Drive (/Real_Dataset/Static/).
# Test our method on two-exposure data. Download static_RGB_data_2exp_rand_motion_release.tgz and unzip to data/
# Results can be found in data/models/CoarseToFine_2Exp/
python run_model.py --gpu_ids 0 --model hdr2E_flow2s_model \
--benchmark real_benchmark_dataset --bm_dir data/static_RGB_data_2exp_rand_motion_release --test_scene all \
--mnet_name weight_net --mnet_checkp data/models/CoarseToFine_2Exp/weight_net.pth --fnet_checkp data/models/CoarseToFine_2Exp/flow_net.pth --mnet2_checkp data/models/CoarseToFine_2Exp/refine_net.pth
# Test our method on three-exposure data. Download static_RGB_data_3exp_rand_motion_release.tgz and unzip to data/
# The results can be found in data/models/CoarseToFine_3Exp/
python run_model.py --gpu_ids 0 --model hdr3E_flow2s_model \
--benchmark real_benchmark_dataset --bm_dir data/static_RGB_data_3exp_rand_motion_release --test_scene all \
--mnet_name weight_net --mnet_checkp data/models/CoarseToFine_3Exp/weight_net.pth --fnet_checkp data/models/CoarseToFine_3Exp/flow_net.pth --mnet2_checkp data/models/CoarseToFine_3Exp/refine_net.pth
Testing on the captured dynamic with GT dataset
The dynamic with GT dataset can be found in Google Drive (/Real_Dataset/Dynamic/).
# Test our method on two-exposure data. Download dynamic_RGB_data_2exp_release.tgz and unzip to data/
python run_model.py --gpu_ids 0 --model hdr2E_flow2s_model \
--benchmark real_benchmark_dataset --bm_dir data/dynamic_RGB_data_2exp_release --test_scene all \
--mnet_name weight_net --fnet_checkp data/models/CoarseToFine_2Exp/flow_net.pth --mnet_checkp data/models/CoarseToFine_2Exp/weight_net.pth --mnet2_checkp data/models/CoarseToFine_2Exp/refine_net.pth
# Test our method on three-exposure data. Download dynamic_RGB_data_3exp_release.tgz and unzip to data/
python run_model.py --gpu_ids 0 --model hdr3E_flow2s_model \
--benchmark real_benchmark_dataset --bm_dir data/dynamic_RGB_data_3exp_release --test_scene all \
--mnet_name weight_net --fnet_checkp data/models/CoarseToFine_3Exp/flow_net.pth --mnet_checkp data/models/CoarseToFine_3Exp/weight_net.pth --mnet2_checkp data/models/CoarseToFine_3Exp/refine_net.pth
Testing on the customized dataset
You have two options to test our method on your dataset. In the first option, you have to implement a customized Dataset class to load your data, which should not be difficult. Please refer to datasets/tog13_online_align_dataset.py.
If you don't want to implement your own Dataset class, you may reuse datasets/tog13_online_align_dataset.py. However, you have to first arrange your dataset similar to TOG13 dataset. Then you can run utils/compute_nbr_trans_for_video.py to compute the similarity transformation matrices between neighboring frames to enable global alignment.
# Use gamma curve if you do not know the camera response function
python utils/compute_nb_transformation_video.py --in_dir /path/to/your/dataset/ --crf gamma --scene_list your_scene_list
HDR evaluation metrics
We evaluate PSRN, HDR-VDP, HDR-VQM metrics using the Matlab code. Please first install HDR Toolbox to read HDR. Then set the paths of the ground-truth HDR and the estimated HDR in matlab/config_eval.m. Last, run main_eval.m in the Matlab console in the directory of matlab/.
main_eval(2, 'Ours')
main_eval(3, 'Ours')
Tonemapping
All visual results in the experiment are tonemapped using Reinhard et al.’s method. Please first install luminance-hdr-cli. In Ubuntu, you may use sudo apt-get install -y luminance-hdr to install it. Then you can use the following command to produce the tonemmapped results.
python utils/tonemapper.py -i /path/to/HDR/
Precomputed Results
The precomputed results can be found in Google Drive (/Results) (TODO).
Training
The training process is described in docs/training.md.
License
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
Citation
If you find this code useful in your research, please consider citing:
@article{chen2021hdr,
title={{HDR} Video Reconstruction: A Coarse-to-fine Network and A Real-world Benchmark Dataset},
author={Chen, Guanying and Chen, Chaofeng and Guo, Shi and Liang, Zhetong and Wong, Kwan-Yee K and Zhang, Lei},
journal=ICCV,
year={2021}
}
2 Feb 04, 2022
4.4k Dec 27, 2022
130 Jan 06, 2023
49 Dec 01, 2022
583 Dec 30, 2022
4 Oct 09, 2021
23 Oct 19, 2022
64 Dec 24, 2022
5 May 24, 2022
45 Dec 15, 2022
2k Dec 29, 2022
76 Dec 11, 2022
2 Dec 09, 2022
20 Feb 03, 2022
31 Dec 11, 2022
13 Sep 23, 2022
27 Oct 14, 2022
5.2k Jan 05, 2023
3 Sep 18, 2021
48 Sep 19, 2022