Modeling Indirect Illumination for Inverse Rendering

Project Page | Paper | Data

Preparation

• Set up the python environment

conda create -n invrender python=3.7
conda activate invrender

pip install -r requirement.txt

• Download our example synthetic dataset from Google Drive

Run the code

Training

Taking the scene hotdog as an example, the training process is as follows.

1. Optimize geometry and outgoing radiance field from multi-view images. (Same as IDR)

   cd code
   python training/exp_runner.py --conf confs_sg/default.conf \
                                 --data_split_dir ../Synthetic4Relight/hotdog \
                                 --expname hotdog \
                                 --trainstage IDR \
                                 --gpu 1

2. Draw sample rays above surface points to train the indirect illumination and visibility MLP.

   python training/exp_runner.py --conf confs_sg/default.conf \
                                 --data_split_dir ../Synthetic4Relight/hotdog \
                                 --expname hotdog \
                                 --trainstage Illum \
                                 --gpu 1

3. Jointly optimize diffuse albedo, roughness and direct illumination.

   python training/exp_runner.py --conf confs_sg/default.conf \
                                 --data_split_dir ../Synthetic4Relight/hotdog \
                                 --expname hotdog \
                                 --trainstage Material \
                                 --gpu 1

Relighting

• Generate videos under novel illumination.

  python scripts/relight.py --conf confs_sg/default.conf \
                            --data_split_dir ../Synthetic4Relight/hotdog \
                            --expname hotdog \
                            --timestamp latest \
                            --gpu 1

Citation

@inproceedings{zhang2022invrender,
  title={Modeling Indirect Illumination for Inverse Rendering},
  author={Zhang, Yuanqing and Sun, Jiaming and He, Xingyi and Fu, Huan and Jia, Rongfei and Zhou, Xiaowei},
  booktitle={CVPR},
  year={2022}
}

Acknowledgements: part of our code is inherited from IDR and PhySG. We are grateful to the authors for releasing their code.