Deep-Relighting

Code for paper "Deep Image-Based Relighting from Optimal Sparse Samples". For more details, please refer to http://viscomp.ucsd.edu/projects/SIG18Relighting

Disclaimer

This software and related data are published for academic and non-commercial use only.

Data

The traing and testing data can be downloaded from: https://drive.google.com/drive/folders/1bmJREP58WNvq9o5nIwNQEEce7cb3uEDk?usp=sharing

"data/orgTrainingImages" contians the original 512x512 training images rendered using mitsuba.
"data/train" contains the cropped 128x128 training patches we use for joint training and refinement.
"data/test" contains the testing dataset rendered using mitsuba.
"data/real" contains the real data we captured at UC San Diego.
"shapes" contains the procedually generated shapes, we created to render the dataset.
"trained/dirs" contains our learnt optimal light directions for different settings.
"trained/joint" contains our trained networks of different training settings after jointly training.
"trained/refine" contains our trained networks of different training settings after refining the Relight Net.
"env" contains a testing environment map.

Platform

This code is based on python 2.7 with tensorflow 1.3.0. It has been tested on Ubuntu 16.04.

Usage

To train the Sample Net and Relight Net jointly with a setting of 90 degree cone and 5 samples. Run:

python run_train_joint.py 90 5 -i ../data/train/joint_npy -o OUTFOLDER

To refine training a Relight Net with our cropped data. Run:

python run_train_refine.py 90 5 -i ../data/train -o OUTFOLDER -j ../trained/joint -d ../trained/dirs

To test a trained Relight Net on our testing dataset. Run:

python run_test.py 90 5 -i ../data/test/mapper_100_sameDir -o OUTFOLDER -w ../trained/refine -d ../trained/dirs

To test a trained Relight Net on one of our real data. Run:

python run_test.py 90 5 -i ../data/real/0 -o OUTFOLDER -w ../trained/refine -d ../trained/dirs

To render one data under a environment map. Run:

python run_renderEnv.py 90 5 -i ../data/real/0 -o OUTFOLDER -w ../trained/refine -d ../trained/dirs

To play with creating new shapes. Run:

python run_genShapes.py -o OUTFOLDER

For detail settings of the programs, please run them with --help or check the source codes.

Citation

If you find this work useful for your research, please cite:

@article{xu2018deep,
  title={Deep image-based relighting from optimal sparse samples},
  author={Xu, Zexiang and Sunkavalli, Kalyan and Hadap, Sunil and Ramamoorthi, Ravi},
  journal={ACM Transactions on Graphics (TOG)},
  volume={37},
  number={4},
  pages={126},
  year={2018},
  publisher={ACM}
}

Contact

Feal free to contact me if there is any question ([email protected])

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

zexiangxu / Deep-Relighting