Source code and data for the paper
Deep Unrolled Low-Rank Tensor Completion for High Dynamic Range Imaging
Truong Thanh Nhat Mai, Edmund Y. Lam, and Chul Lee
IEEE Transactions on Image Processing, vol. 31, pp. 5774-5787, 2022
https://doi.org/10.1109/TIP.2022.3201708

We also provide source code for the matrix completion version published in ICIP for references
Ghost-Free HDR Imaging Via Unrolling Low-Rank Matrix Completion
Truong Thanh Nhat Mai, Edmund Y. Lam, and Chul Lee
2021 IEEE International Conference on Image Processing (ICIP)
https://doi.org/10.1109/ICIP42928.2021.9506201

For PDFs, please visit https://mtntruong.github.io/

If you have any question, please open an issue.
The algorithm can also be applied to other applications. Please feel free to ask if you need help with training the algorithm using other datasets, either HDR imaging or other applications.

The proposed algorithm is implemented in Python using PyTorch 1.11.

Download samples from Microsoft OneDrive

The folder contains four ZIP files:

• Training_Samples.zip: 13000 training samples
• Training_Samples_ICIP.zip: 13000 training samples used in the ICIP paper
• HDM-HDR_Test_Samples.zip: Warped exposures of the HDM-HDR dataset
• HDRv_Test_Samples.zip: Warped exposures of the HDRv dataset

If you do not have time to retrain the network, you may use pretrained weights
Download pretrained weights from Microsoft OneDrive

The folder contains two PTH files:

• LRT-HDR_net.pth: pretrained weight of LRT-HDR
• ICIP_net.pth: pretrained weight of the matrix completion network (ICIP paper)

Please use env.yml to create an environment in Anaconda

conda env create -f env.yml

Then activate the environment

conda activate lrt

If you want to change the environment name, edit the first line of env.yml before creating the environment.

Extract Training_Samples.zip to obtain the folder Training_Samples, then run

python train_auto.py --data_path=/path/to/Training_Samples
# or
python train_manual.py --data_path=/path/to/Training_Samples

While train_auto.py adjusts learning rate automatically, it usually yields worse performance (still better than competing algorithms). Using train_manual.py provides best results but you have to manually adjust learning rate. I have tried several ways to update learning rate during training, including torch.optim.lr_scheduler, but manually adjusting learning rate is always better.

When using train_manual.py, please cancel the training process every 10 epochs then rerun to update the learning rate using the following commands

# After 10th epoch
python train_manual.py --data_path=/path/to/Training_Samples --resume=./checkpoints/epoch_10.pth --set_lr=1e-6
# After 20th epoch
python train_manual.py --data_path=/path/to/Training_Samples --resume=./checkpoints/epoch_20.pth --set_lr=1e-7
# After 30th epoch
python train_manual.py --data_path=/path/to/Training_Samples --resume=./checkpoints/epoch_30.pth --set_lr=1e-8
# Stop after 40th epoch and you are done

After the training process completes, you should use the weight named epoch_40.pth for testing.

Extract HDM-HDR_Test_Samples.zip and HDRv_Test_Samples.zip to obtain the preprocessed test sets of the HDM-HDR and HDRv datasets, respectively, then run

# For HDM-HDR dataset
python test.py --data=hdm --data-path=/path/to/HDM-HDR_Test_Samples --output-path=./HDM-HDR_results --checkpoint=/path/to/LRT-HDR_net.pth
# For HDRv dataset
python test.py --data=hdrv --data-path=/path/to/HDRv_Test_Samples --output-path=./HDRv_results --checkpoint=/path/to/LRT-HDR_net.pth

The synthesized HDR images will be written in the folder specified by --output-path

Download dataset and results from Microsoft OneDrive

The folder contains two ZIP files:

• Datasets.zip: This file contains 187 and 32 multi-exposure image sets generated from the HDM-HDR and HDRv datasets, respectively, as described in the paper.
• All_Synthesized_Results.zip: We also provide HDR images synthesized by the proposed algorithm and all other competing algorithms, so that you can inspect the results to your heart's content without rerunning 10 algorithms.

As you may be aware, our dataset and that of NTIRE challenge are all generated from the videos of HDM-HDR dataset. However, the data formats of the generated LDR and HDR images are different. Our dataset has the same format as the Kalantari and Ramamoorthi's dataset, that means it is fully compatible with existing HDR algorithms that are designed for Kalantari and Ramamoorthi's dataset. We also generate an additional test set from HDRv with the same format.

If our research or dataset are useful for your research, please kindly cite our work

@article{Mai2022,
    author={Mai, Truong Thanh Nhat and Lam, Edmund Y. and Lee, Chul},
    journal={IEEE Transactions on Image Processing}, 
    title={Deep Unrolled Low-Rank Tensor Completion for High Dynamic Range Imaging}, 
    year={2022},
    volume={31},
    number={},
    pages={5774-5787},
    doi={10.1109/TIP.2022.3201708}
}

or if you prefer the low-rank matrix completion algorithm

@inproceedings{Mai2021,
    author={Mai, Truong Thanh Nhat and Lam, Edmund Y. and Lee, Chul},
    booktitle={2021 IEEE International Conference on Image Processing (ICIP)},
    title={Ghost-Free HDR Imaging Via Unrolling Low-Rank Matrix Completion},
    year={2021},
    volume={},
    number={},
    pages={2928-2932},
    doi={10.1109/ICIP42928.2021.9506201}
}

Also, citing the original HDM-HDR and HDRv video datasets is appreciated.