deep-high-dynamic-range

Tensorflow implementation of SIGGRAPH 17 paper: Deep High Dynamic Range Imaging of Dynamic Scenes

Update 2021-05-26: This implementation is out of date and I do not intend to maintain it as it's only a course project during my undergrad, so I've archived it.

Installation

This implementation requires python3, tensorflow 2.0 and opencv. Please install dependencies via:

pip install tensorflow==2.0
pip install opencv-python
pip install opencv-contrib-python

Tested on MacOS 10.15 and CentOS 7.0

Training

First download the dataset

cd data/
./download.sh

Then run preprocess.py. This file accepts train or test as optional argument to generate only train/test set. The raw data will be transformed into tfrecords format and stored in tf-data folder.

python preprocess.py

Finally, run train.py, this file accepts a argument specifying model type: direct, we or wie

python train.py [model_type]

Testing

Use pretrained weights for testing, run test.py. This file again accepts a model type string and an additional argument specifying checkpoint path.

python test.py [model_type] [checkpoint_path]

Example:

python test.py direct saved-checkpoints/deepflow-direct/model.ckpt-100

Reference

1. Kalantari, N.K., Ramamoorthi, R.: Deep High Dynamic Range Imaging of Dynamic Scenes. ACM TOG 36(4) (2017)