PhySG: Inverse Rendering with Spherical Gaussians for Physics-based Relighting and Material Editing

• CVPR 2021. Project page: https://kai-46.github.io/PhySG-website/

Quick start

• Create conda environment

  conda env create -f environment.yml
  conda activate PhySG

• Download example data from google drive.

• Optimize for geometry and material given a set of posed images and object segmentation masks

  cd code
  python training/exp_runner.py --conf confs_sg/default.conf \
                            --data_split_dir ../example_data/kitty/train \
                            --expname kitty \
                            --nepoch 2000 --max_niter 200001 \
                            --gamma 1.0

• Render novel views, relighting and mesh extraction, etc.

  cd code
  # use same lighting as training
  python evaluation/eval.py --conf confs_sg/default.conf \
                            --data_split_dir ../example_data/kitty/test \
                            --expname kitty \
                            --gamma 1.0 --resolution 256 --save_exr
  # plug in new lighting                              
  python evaluation/eval.py --conf confs_sg/default.conf \
                            --data_split_dir ../example_data/kitty/test \
                            --expname kitty \
                            --gamma 1.0 --resolution 256 --save_exr \
                            --light_sg ./envmaps/envmap3_sg_fit/tmp_lgtSGs_100.npy

Tips: for viewing exr images, you can use tev hdr viewer.

Some important pointers

• code/model/sg_render.py: core of the appearance modelling that evaluates rendering equation using spherical Gaussians.
  • code/model/sg_envmap_convention.png: coordinate system convention for the envmap.
• code/model/sg_envmap_material.py: optimizable parameters for the material part.
• code/model/implicit_differentiable_renderer.py: optimizable parameters for the geometry part; it also contains our foward rendering code.
• code/training/idr_train.py: SGD optimization of unknown geometry and material.
• code/evaluation/eval.py: novel view rendering, relighting, mesh extraction, etc.
• code/envmaps/fit_envmap_with_sg.py: represent an envmap with mixture of spherical Gaussians. We provide three envmaps represented by spherical Gaussians optimized via this script in the 'code/envmaps' folder.

Prepare your own data

• Organize the images and masks in the same way as the provided data.
• Make sure object of interest is inside the unit sphere by properly normalizing your camera parameters.
• As to camera parameters, we follow the same convention as NeRF++ to use OpenCV conventions.

Acknowledgements: this codebase borrows a lot from the awesome IDR work; we thank the authors for releasing their code.