Neural BRDF

Code repository for the paper:

Neural BRDF Representation and Importance Sampling
Alejandro Sztrajman, Gilles Rainer, Tobias Ritschel, Tim Weyrich
Computer Graphics Forum (CGF), 40(6), pp. 332-346, 2021 (Oral Presentation at EGSR 2022).

Project Page | Paper

Content

binary_to_nbrdf/
  binary_to_nbrdf.py: python script to encode one or more materials (in MERL binary format) as NBRDF neural networks.
    - Usage: ./binary_to_nbrdf.py <material.binary> (optionally specify multiple materials)
    - NBRDF is written as a keras .h5 network (<material>.h5, <material>.json). Sample pre-trained networks can be downloaded from the project webpage.
    - Generated NBRDF networks are 6 x 21 x 21 x 3 (675 weights).
    - Script can be directly used on any BRDF in binary format (materials can be downloaded from MERL or from Nielsen et al. 2015).
    - Code has been tested with the following module versions:
    -- keras 2.2.5
    -- tensorflow-gpu 1.13.1

  h5_to_npy.py: python script to convert a .h5 NBRDF file from keras into a set of .npy files that can be read from Mitsuba.
    - Usage: ./h5_to_npy.py <material.h5> (optionally specify multiple .h5 files)
    - The script creates a folder "npy" in the location of <material.h5>, with 6 files representing the NBRDF weights.

  pytorch_code/
    NBRDF training script. Alternative implementation in PyTorch by Michael Fischer.

data/
  merl_nbrdf.zip: pre-trained NBRDF models of all materials from the MERL database.
  rgl-isotropic_nbrdf.zip: pre-trained NBRDF models of isotropic materials from the RGL database.
  nielsen_nbrdf.zip: pre-trained NBRDF models of materials from Nielsen et al. (2015).

mitsuba/
  C++ codes required to render NBRDF materials in the Mitsuba renderer (https://www.mitsuba-renderer.org/index_old.html).
  Installation:
    - Copy all files from mitsuba/bsdfs into $MITSUBA/src/bsdfs/ ($MITSUBA=Mitsuba installation folder).
    - Edit the file $MITSUBA/src/bsdfs/SConscript and add the following line:
      plugins += env.SharedLibrary('nbrdf_npy', ['nbrdf_npy.cpp'])
    - (Re)compile Mitsuba
  Usage:
    - sample_scene.xml is a Mitsuba scene ready to render using the material "blue-acrylic" in "data/merl_nbrdf/npy/".
    - Simply run: mitsuba sample_scene.xml. This will generate an output rendering sample_scene.exr.
    - To replace the material simply change the definition of the string "nn_basename", but remember to run h5_to_npy.py on the desired .h5 file first.

Usage Summary

1) Follow installation instructions for mitsuba files
2) In the project webpage there are pre-trained NBRDFs for materials from multiple databases. Run h5_to_npy.py on the desired NBRDF material (.h5 format).
3) Edit sample_scene.xml and modify the variable "nn_basename" to point to the desired material files.
4) Run: mitsuba sample_scene.xml

BibTeX

If you find our work useful, please cite:

@article{sztrajman2021nbrdf,
    title = {Neural {BRDF} Representation and Importance Sampling},
    author = {Sztrajman, Alejandro and Rainer, Gilles and Ritschel, Tobias and Weyrich, Tim},
    journal = {Computer Graphics Forum},
    volume= 40,
    number= 6,
    pages = {332--346},
    month = sep,
    year = 2021,
    doi = {https://doi.org/10.1111/cgf.14335},
    url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.14335}
}

Contact

If you have any questions, please email Alejandro Sztrajman at a.sztrajman@ucl.ac.uk.

asztr / Neural-BRDF