Look-Ahead Training with Learned Reflectance Loss for Single-Image SVBRDF Estimation

This is code of "Look-Ahead Training with Learned Reflectance Loss for Single-Image SVBRDF Estimation" Project | Paper

Set up environment

To set up environment, please run this command below (tesed in Linux environment):

conda env create -f env.yml

Inference

Before running inference, please download:

1. our pretrained model from this link
2. centeralized MaterialGAN and our dataset with ground truth link

please save the download model to ./ckpt/ and extract data to ./dataset:

To run inference on our and MaterialGAN's dataset with ground truth, please use this command:

python meta_test.py --fea all_N1 --wN_outer 80 --gamma --cuda --test_img $mode --name $name --val_step 7 --wR_outer 5 --loss_after1 TD --Wfea_vgg 5e-2 --Wdren_outer 10 --WTDren_outer 10 --adjust_light

where $mode set as OurReal2 for our test dataset and MGReal2 for MaterialGAN dataset, $name represents the saved path

To run inference on real captured dataset without ground truth, please first centeralized the specular highlight of input image and then run this command:

python meta_test.py --val_root $path ---fea all_N1 --wN_outer 80 --gamma --cuda --test_img Real --name $name --val_step 7 --wR_outer 5 --loss_after1 TD --Wfea_vgg 5e-2 --Wdren_outer 10 --WTDren_outer 10 --adjust_light

where $path point to the directory of test real images, $name represents the saved path. The final feature maps are saved to $name\fea and optimization process at step 0,1,2,5,7 are saved to $name\pro

Our Dataset

We also provide higher resolution version of unscaled real scenes: link

Citation

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

@article{zhou2022look,
  title={Look-Ahead Training with Learned Reflectance Loss for Single-Image SVBRDF Estimation},
  author={Zhou, Xilong and Kalantari, Nima Khademi},
  journal={ACM Transactions on Graphics (TOG)},
  volume={41},
  number={6},
  pages={1--12},
  year={2022},
  publisher={ACM New York, NY, USA}
}

Contact

This code is not clean version, will clean it up soon. feel free to email me if you have any questions: [email protected]. Thanks for your understanding!