On cross-validation and resampling of BRDF data measurements

Vlastimil Havran, Attila Neumann, Georg Zotti, Werner Purgathofer, Hans-Peter SeidelSCCG 2005 · DOI

TL;DR

This paper tackles the challenge of validating surface reflectance data (BRDF) from different measurement setups, especially when calibration is unknown or inconsistent. They introduce a novel resampling parameterization and a colorimetric comparison method to bridge the gap between uncalibrated measurements, demonstrating its effectiveness on real-world materials.

What Problem Does This Paper Address?

Accurate surface reflectance properties, known as Bidirectional Reflectance Distribution Functions (BRDFs), are crucial for realistic computer graphics rendering. However, obtaining and comparing BRDF data from different measurement systems is fraught with difficulties. These challenges arise from variations in equipment, calibration procedures (or lack thereof), and data representation. This paper specifically addresses the problem of cross-validating BRDF data acquired from two distinct measurement setups: one from a commercial company (Integra) and another developed within an EU project (UBO). The key issue is that the UBO setup's data is uncalibrated, meaning it has unknown multiplicative factors, and the accuracy of the Integra data is also not fully specified. Direct comparison is impossible without addressing these discrepancies, making validation a significant hurdle.

Methods

The authors propose a multi-pronged approach to tackle the cross-validation problem:

  • BRDF Averaging: For the UBO setup, which captures Bidirectional Texture Functions (BTFs), the first step is to average the BTF data over spatial locations to obtain BRDF values. This is a linear operation performed on HDR images.
  • Color Space Transformation: Both the Integra spectral reflectance data and the UBO RGB data are transformed into the CIE XYZ color space. Subsequently, they are converted to the perceptually uniform CIE 1976 L*a*b* color space for more meaningful color difference comparisons.
  • Novel BRDF Resampling Parameterization: A key contribution is a new parameterization for resampling BRDF data. This parameterization is designed to be bicontinuous and aligns the data along the direction of specular reflection. This is crucial for accurately interpolating and comparing data, especially for glossy materials where reflectance changes rapidly in that direction. It maps a 2D unit square to a 3D sphere with a single point of discontinuity, ensuring better interpolation properties.
  • Reciprocity Check: As a fundamental validation step, the Helmholtz reciprocity principle is applied. This principle states that the BRDF should be the same when the incident and outgoing directions are swapped. This check helps identify positioning errors or fundamental issues in the measurement process.
  • Colorimetric Comparison: To compare the uncalibrated data, the authors develop a method to estimate a color transformation matrix. This matrix maps color values derived from the reference BRDF data to the color values obtained from the UBO BTF measurements. This is achieved by minimizing the difference between the two color sets using a least-squares fitting method.

Datasets

The paper uses four material samples for validation: "ceiling," "floortile," "pinkwall," and "walkway." These samples were measured by two different setups:

  • Integra Data: Acquired by a commercial company in 2001, this data is described as spectral reflectance data.
  • UBO Data: Acquired by the UBO setup, this data consists of BTF images captured in linear RGB color space (Radiance HDR format).

The paper does not introduce new benchmark datasets but rather uses existing measurements from two distinct sources for cross-validation.

Key Insights and Ideas

  • Bridging the Uncalibrated Gap: The core insight is that even with uncalibrated data, meaningful cross-validation is possible by leveraging fundamental physical principles (reciprocity) and perceptual color spaces.
  • The Power of Perceptual Color Spaces: Using CIE 1976 L*a*b* for color difference evaluation provides a more intuitive measure of similarity than raw RGB or XYZ values, aligning better with human perception.
  • Resampling Matters: The development of a novel, bicontinuous parameterization for resampling BRDF data is critical. It addresses the discontinuity issues of other parameterizations and improves interpolation accuracy, especially for specular materials.
  • Reciprocity as a Foundation: The reciprocity check serves as a robust initial validation step, capable of detecting fundamental measurement errors even before complex cross-comparison.

Key Results

The paper demonstrates the successful cross-validation of BRDF data from two different measurement setups.

  • Reciprocity Check: The reciprocity check was applied to single datasets and showed that the measurements largely fulfilled the principle, indicating good overall data quality.
  • Color Comparison: The mean color differences (DE and DEexp) between the transformed Integra data and the UBO data were found to be generally small, with values often below 1 (indicating "barely detectable" differences). While some materials (like "walkway") showed larger maximum differences, these were attributed to factors like self-shadowing and the inherent limitations of interpolating highly specular data. The authors conclude that the resulting differences are within the expected range of uncertainties.

Limitations and Future Work

  • Limitations:
    • Polarization: The measurements did not account for polarization effects, which could influence reflectance properties.
  • Integra Accuracy: The exact accuracy and measurement uncertainty of the Integra setup were not provided, making a definitive quantitative assessment of the validation challenging.
  • Border Effects: Averaging BTF texels to obtain BRDF can introduce deviations due to border effects.
  • Future Work:
    • Direct BTF Validation: The authors suggest that direct validation of BTF data, rather than just averaged BRDF, would be a valuable future direction.
  • Information Theory: Exploring the use of tools from information theory, such as mutual entropy, for cross-validation of BRDF measurements is proposed as an interesting avenue.
  • Analytical BTF Models: Direct validation of analytical BTF models is also highlighted as a desirable future research topic.

Summary updated

BibTeX
@inproceedings{havran2005cross,
  author = {Vlastimil Havran and Attila Neumann and Georg Zotti and Werner Purgathofer and Hans-Peter Seidel},
  title = {{On cross-validation and resampling of BRDF data measurements}},
  booktitle = {SCCG},
  year = {2005},
  doi = {10.1145/1090122.1090149},
}
Figure