Uncertainty for SVBRDF Acquisition using Frequency Analysis

Code release for SIGGRAPH 2025 paper "Uncertainty for SVBRDF Acquisition using Frequency Analysis"
Ruben Wiersma, Julien Philip, Milos Hasan, Krishna Mullia, Fujun Luan, Elmar Eisemann, Valentin Deschaintre

Setup • Quickstart • Instructions • License • Roadmap • Citation

We quantify uncertainty for SVBRDF acquisition from multi-view captures (left) using entropy (right). We significantly accelerate the, otherwise heavy, computation in the frequency domain (Spherical Harmonics), yielding a practical, efficient method.

🗝 Setup

^{▴ Back to top}

Environment

To run our code, you need to setup a (conda) environment and install the svbrdf_uncertainty package. First, clone the repository:

git clone https://github.com/rubenwiersma/svbrdf_uncertainty.git

Then move to the main directory, create an environment with python 3.10 and install the svbrdf_uncertainty package:

cd svbrdf_uncertainty
conda create -n svbrdf_uncertainty python=3.10 -y
conda activate svbrdf_uncertainty
pip install .

This will also install the required dependencies, such as Mitsuba, nvdiffrast, and torch.

Make sure the environment is activated when you run the experiments.

conda activate svbrdf_uncertainty

🚀 Quickstart

^{▴ Back to top}

To check if the setup was done correctly, you can run our optimization scripts for the spherical harmonics model and Mitsuba on the plane toy example:

python opt_sh.py # Runs our method
python opt_mi.py # Runs Mitsuba

This runs the optimization and entropy computation for our method. The results will be placed in the folder out/plane/[sh/mitsuba]/default.

The results folders contain .exr files with the output textures (base_color.exr, entropy.exr - uncertainty, metallic.exr, roughness.exr) and an overview figure brdf_overview.png with these textures and re-renders.

🪄 Instructions

^{▴ Back to top}

To run the experiments in our paper, specifically for Stanford ORB, you need to download Stanford ORB and run the corresponding scripts.

Skip to:
Data • Running experiments • Configurations

Stanford ORB Data

Download and uncompress the files for Stanford ORB using the following script:

source stanford_orb_download.sh

This script will create a new folder outside of the repository in ../data/Stanford ORB and download Stanford ORB to this folder.

Finally, run the following python script to correct the normals in the .obj files given by Stanford ORB (their y and z components are swapped).

conda activate svbrdf_uncertainty
python scripts/normal_correction_stanford_orb.py

Note: This script and many other locations in the code assume that you downloaded Stanford ORB to ../data/Stanford ORB.

Running the experiments

You can use the following scripts to run the Stanford ORB benchmark for our approach and Mitsuba:

python scripts/opt_sh_stanford_orb.py # Our method
python scripts/opt_mi_stanford_orb.py # Mitsuba

These scripts estimate material textures for each scene in Stanford ORB, perform relighting (combined ~10-15min), and compute the metrics in the Stanford ORB benchmark (~4min).

You need to set the path to the dataset in opt_[sh/mi]_stanford_orb.py on line 15/16 if you downloaded Stanford ORB to a different directory.

The benchmark results will be stored in out/stanford_orb/sh_baseline.json and out/stanford_orb/mi_baseline.json. You can format the results for easier readibility to LaTeX using the script scripts/json_to_latex.py:

python scripts/json_to_latex.py out/stanford_orb

The output is a .tex file in out/stanford_orb/stanford_orb.tex, containing the results from all .json files in out/stanford_orb. These tables correspond to the results in Table 2 of the paper.

You can also inspect the resulting textures and entropy maps in the folders for each scene, e.g., out/stanford_orb/blocks_scene002.

Configurations

We make extensive use of the gin config package to configure experiments. Functions decorated with @gin.configurable can be configured with .gin files. We use this in the following places:

• opt_sh.py, optimize_material() - The main optimization function. An example configuration script is provided in experiments/scenes/plane.
• scripts/opt_sh_stanford_orb.py, run_benchmark() - The benchmark function. We use this to run different variants of our method (e.g., power spectrum vs. angular domain) on the entire benchmark. The benchmark script takes a folder as input (in experiments/configs) and will recursively look through all its sub-folders to find config files. If you want to run different variants of the benchmark, simply create a .gin file in a folder and pass the folder as an argument to opt_sh_stanford_orb.py:

python scripts/opt_sh_stanford_orb.py experiments/configs/new_folder

🔖 License

^{▴ Back to top}

This repository is licensed under the MIT License.

🗺 Roadmap

^{▴ Back to top}

We aim to release code for the synthetic benchmark. This might take a bit longer due to licensing constraints.

📖 Citation

^{▴ Back to top}

Please cite our paper if this code contributes to an academic publication:

@inproceedings{wiersma2025svbrdfuncertainty,
    author = {Wiersma, Ruben and Philip, Julien and Hašan, Miloš and Mullia, Krishna and Luan, Fujun and Eisemann, Elmar and Deschaintre, Valentin},
    title = {Uncertainty for SVBRDF Acquisition using Frequency Analysis},
    year = {2025},
    isbn = {979-8-4007-1540-2/2025/08},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3721238.3730592},
    doi = {10.1145/3721238.3730592},
    booktitle = {SIGGRAPH Conference Papers '25},
    location = {Vancouver, BC, CA},
    series = {SIGGRAPH Conference Papers '25}
}