This repository describes the training and testing datasets used in HDR Video Reconstruction: A Coarse-to-fine Network and A Real-world Benchmark Dataset, ICCV 2021. We also provide the python scripts used for processing the raw data and capturing real data. Python scripts in this repository work with python 3.5+.
Datasets can be downloaded in BaiduYun. Thumbnails for the captured dataset can be found in this page.
- Cinematic Wide Gamut HDR video: https://www.hdm-stuttgart.de/vmlab/hdm-hdr-2014/.
- LiU HDRv Repository - Resources: http://www.hdrv.org/Resources.php.
A copy of the above two datasets can be found in BaiduYun (Online_HDR_Dataset/) in case you cannot download them. Please acknowledge the corresponding papers if you use them.
- Kalantari13 dataset: this dataset consists of 9 videos for dynamic scenes in RGB image format: https://www.ece.ucsb.edu/~psen/PaperPages/HDRVideo/.
We used this dataset for qualitative evaluation. The prepared dataset can be found in BaiduYun (Real_Dataset/TOG13_Dynamic_Dataset.tgz).
Directory structure
TOG13_Dynamic_Dataset/
├── 2Exp_scenes.txt # list of 2-exposure scenes
├── 3Exp_scenes.txt # list of 3-exposure scenes
├── BaslerCRF.mat # camera response function
├── CheckingEmail-3Exp-2Stop/
│ ├── Affine_Trans_Matrices/ # precomputed similarity transformation between neighboring frames
│ ├── Exposures.txt # exposure information
│ ├── HDR/ # HDR estimation of Kalantari13's method
│ ├── HDR_TOG13_PyTM/ # tonemapped results Kalantari13's method
│ ├── img_hdr_list.txt # list of paired image and hdr
│ ├── img_list.txt # list of images
│ └── TMInfo.txt # parameters for Reinhard et al.'s method
├── Cleaning-3Exp-2Stop/
│ └── ...
├── Dog-3Exp-2Stop/
│ └── ...
├── Fire-2Exp-3Stop/
│ └── ...
├── Ninja-2Exp-3Stop/
│ └── ...
├── ThrowingTowel-2Exp-3Stop/
│ └── ...
├── TMinfo.txt
└── WavingHands-2Exp-3Stop/
└── ...Since there is no publicly available real video dataset with alternating exposures and their ground-truth HDR, we resort to synthetic data for training.
We generate the synthetic training dataset using 21 videos from Cinematic Wide Gamut HDR-video (list of 13 videos) and LiU HDRv Repository - Resources (list of 8 videos). To generate the synthetic training data, we first cropped 7-frame HDR patches and save them as 16bit numpy arrays to hard disk. During training, We re-expose the HDR images into LDR images in an on-line manner. The command for cropping patches is as follows:
# please create a scenes.txt file that stores the scene list in /path/to/HDR/videos/.
python gen_syn_train_dataset.py --in_dir /path/to/HDR/videos/The prepared datasets can be downloaded in BaiduYun (Synthetic_Dataset/).
We have also used the Vimeo-90K dataset for training, please download it from http://toflow.csail.mit.edu/ and unzip it.
Directory structure
Synthetic_Train_Data_LiU_HDRv
├── Images
│ ├── bridge
│ ├── bridge_2
│ ├── hallway
│ ├── hallway2
│ ├── river
│ ├── students
│ ├── water
│ └── window
├── list.txt
├── train_list.txt
└── val_list.txt
Synthetic_Train_Data_HdM-HDR-2014
├── Images
│ ├── bistro_01
│ ├── bistro_02
│ ├── bistro_03
│ ├── cars_closeshot
│ ├── cars_fullshot
│ ├── cars_longshot
│ ├── fireplace_01
│ ├── fireplace_02
│ ├── hdr_testimage
│ ├── showgirl_01
│ ├── showgirl_02
│ ├── smith_hammering
│ └── smith_welding
├── list.txt
├── train_list.txt
└── val_list.txt
The synthetic test dataset was generated using two HDR videos (i.e., POKER FULLSHOT and CAROUSEL FIREWORKS) from Cinematic Wide Gamut HDR-video, which are not used for training. Each video contains 60 frames and has a resolution of 1920 × 1080. Random Gaussian noise was added to the low-exposure images.
The source HDR videos can be found in BaiduYun (Synthetic_Dataset/HDR_Source_Videos_for_Sythetic_Testset.tgz). To generate videos with alternating exposures, you can run
# for two alternating exposures
python gen_syn_test_dataset.py --in_dir /path/to/HDR_Source_Videos_for_Sythetic_Testset --expo_n 2
# for three alternating exposures
python gen_syn_test_dataset.py --in_dir /path/to/HDR_Source_Videos_for_Sythetic_Testset --expo_n 3The generated videos with alternating exposures can be found in BaiduYun (Synthetic_Dataset/HDR_Synthetic_Test_Dataset.tgz).
To facilitate a more comprehensive evaluation on real data, we captured a real-world dataset and generated reliable ground truth HDR for evaluation. We used an off-the-shelf Basler acA4095-30uc camera for capturing videos with alternating exposures (i.e., two and three exposures) in a variety of scenes.
Three different types of video data are captured, namely, the static scene with GT, dynamic scenes with GT, and dynamic scenes without GT. The python code for controlling the camera can be found in ./data_capture/.
We saved the raw data of the captured videos and performed demosaicing, white balance, color correction, and gamma compression to convert the raw data to RGB data using the recorded metadata. In this paper, we rescaled the images to 1536 × 813 for evaluation.
This dataset can be found in BaiduYun (Real_Dataset/Static/).
The captured raw datasets are stored in static_raw_data_2exp_release.tgz and static_raw_data_3exp_release.tgz.
To generate the RGB image datasets from the captured raw data, you can use the following command:
# Step 1. Convert Raw to RGB
python cvt_raw_to_rgb_scenes.py --in_dir /path/to/static_raw_data_2exp_release --max_frame_num 6
python cvt_raw_to_rgb_scenes.py --in_dir /path/to/static_raw_data_3exp_release --max_frame_num 9
# The generated RGB dataset are stored in static_RGB_data_2exp_release.tgz and static_RGB_data_3exp_release.tgz
# Step 2. Apply random global motion
python apply_random_motion_static_data.py --in_dir /path/to/static_RGB_data_2exp_release
python apply_random_motion_static_data.py --in_dir /path/to/static_RGB_data_3exp_releaseThe global motion augmented RGB dataset are stored in static_RGB_data_2exp_rand_motion_release.tgz and static_RGB_data_3exp_rand_motion_release.tgz
This dataset can be found in BaiduYun (Real_Dataset/Dynamic/).
Note that we first rearranged the captured raw data according to the description in Section 4 of the paper.
The prepared raw datasets are stored in dynamic_raw_data_2exp_release.tgz and dynamic_raw_data_3exp_release.tgz.
To generate the RGB images from the captured raw data, you can use the command:
# Convert Raw to RGB
python cvt_dynamicGT_raw_to_rgb_scenes.py --in_dir /path/to/dynamic_raw_data_2exp_release
python cvt_dynamicGT_raw_to_rgb_scenes.py --in_dir /path/to/dynamic_raw_data_3exp_releaseThe processed RGB datasets are stored in dynamic_RGB_data_2exp_release.tgz and dynamic_RGB_data_3exp_release.tgz.
This dataset can be found in BaiduYun (Real_Dataset/Dynamic_noGT/).
The captured raw data are stored in dynamic_raw_data_noGT_2exp.tgz and dynamic_raw_data_noGT_3exp.tgz.
Similarly, you can obtain the RGB images using the command:
python cvt_raw_to_rgb_scenes.py --in_dir /path/to/dynamic_raw_data_noGT_2exp --merge_hdr # do not merge hdr
python cvt_raw_to_rgb_scenes.py --in_dir /path/to/dynamic_raw_data_noGT_3exp --merge_hdr # do not merge hdr
The processed RGB datasets in JPG format (due to the large file size, please generate TIF dataset by yourself) are stored in dynamic_data_noGT_2exp_RGB_JPG.tgz and dynamic_data_noGT_3exp_RGB_JPG.tgz.
If you find this dataset or 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}
}