Application of RGB BRDF Synthesis methods to Spectral BRDFs and the applicability of a Fourier spectral representation

Abstract

Spectral path tracing is used to create highly realistic scenes requiring spectral BRDFS, but BRDF synthesis is mostly limited to RGB BRDFs. To allow the creation of new spectral BRDFs this thesis converts interpolation based RGB BRDF synthesis methods to support spectral representations, and their effectiveness is checked using the RGL Material Database. These techniques are based on dimensionality reduction techniques, for this PCA and an Auto Encoder are used. Two different spectral representations are used and compared, one being a base sampled representation and the other a Fourier based representation. Log-relative mapping is often used to improve performance of the dimensionality reduction, since this mapping method does not work on the Fourier representation a new atan based mapping function is proposed and tested. It is shown that the atan-mapped Fourier representation performs poorly under dimensionality reduction due to its sensitivity to noise. Additionally, spectral BRDFs require a significant increase in components to be represented accurately compared to RGB BRDFs, complicating synthesis techniques that depend on a low dimensionality. While the Auto Encoder allows for a better reduction compared to PCA, the dimensionality is still too large for these RGB oriented synthesis techniques.