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
We provide testing and training codes. Details of the training and testing dataset can be found in DeepHDRVideo-Dataset. Datasets, the trained models, and the computed results can be download in Google Drive or BaiduYun (TODO).
This method 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
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/
.
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
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 required 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
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
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
The dynamic with GT dataset can be found in Google Drive (/Real_Dataset/Dynamic_noGT/
).
# Test our method on two-exposure data. Download dynamic_data_noGT_2exp_RGB_JPG.tgz and unzip to data/
# Note that we provide the JPG dataset only for illustrating the testing process
# 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/dynamic_data_noGT_2exp_RGB_JPG --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
# It is similar to test on three-exposure data
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 the 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
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')
All visual results in the experiment are tonemapped using Reinhard et al.โs 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/
The precomputed results can be found in Google Drive (/Results
).
The training process is described in docs/training.md.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
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}
}