Perceptual quality of BRDF approximations: dataset and metrics

Guillaume Lavoué1,2 Nicolas Bonneel1 Jean-Philippe Farrugia1,3 Cyril Soler4
1CNRS, LIRIS, France 2ENISE/Ecole Centrale de Lyon 3Université Lyon1 4INRIA

In Computer Graphics Forum (Proceedings of Eurographics 2021)

Teaser
Several analytical approximations of the gold-metallic-paint3 MERL BRDF using classical models, with mean subjective opinion scores reflecting their perceived material similarity with the tabulated reference (between 1--very poor and 5--excellent).

Abstract

Bidirectional Reflectance Distribution Functions (BRDFs) are pivotal to the perceived realism in image synthesis. While measured BRDF datasets are available, reflectance functions are most of the time approximated by analytical formulas for storage efficiency reasons. These approximations are often obtained by minimizing metrics such as $L_2$---or weighted quadratic---distances, but these metrics do not usually correlate well with perceptual quality when the BRDF is used in a rendering context, which motivates a perceptual study. The contributions of this paper are threefold. First, we perform a large-scale user study to assess the perceptual quality of 2026 BRDF approximations, resulting in 84138 judgments across 1005 unique participants. We explore this dataset and analyze perceptual scores based on material type and illumination. Second, we assess nine analytical BRDF models in their ability to approximate tabulated BRDFs. Third, we assess several image-based and BRDF-based ($L_p$, optimal transport and kernel distance) metrics in their ability to approximate perceptual similarity judgments.

Paper (16MB)
Supplementary Material (37MB)

Dataset

Online dataset exploration - Javascript page allowing to explore our data.
BRDFs - The complete dataset of 2026 BRDFs encoded in both MERL and TITOPO formats.
Rendered images (336MB) - 2796 images (580×900 resolution), 2026 of which were rendered with the Grace lighting, and 770 of which were rendered with the Uffizi lighting.
Raw subjective scores - CSV file containing the 84,138 human judgments, used to compute the statistics.
Mean opinion scores - CSV file containing the 1926 mean opinion scores from the Grace map, used to benchmark the metrics.
Metrics values - CSV file containing the metrics values for the 1926 BRDF approximations.
BRDF classification -  CSV file containing our dielectrics/metals classification of the MERL dataset.
Source code  - Source code for computing the Lp and Optimal Transport metrics, and souce code for conversion between MERL and TITOPO.

Acknowledgements

We are deeply grateful to Derek Nowrouzezahrai and Mahdi M. Bagher for providing the data from their approximation algorithm. We also thank Abir Zendagui for her help to generate the approximations and images.