Acquisition and Representation of Material Appearance for Editing and Rendering (thesis)
Providing computer models that accurately characterize the appearance of a wide class of materials is of great interest to both the computer graphics and computer vision communities. The last ten years has witnessed a surge in techniques for measuring the optical properties of physical materials. As compared to conventional techniques that rely on hand-tuning parametric light reflectance functions, a data-driven approach is better suited for representing complex real-world appearance. However, incorporating these representations into existing rendering algorithms and a practical production pipeline has remained an open research problem.
One common approach has been to fit the parameters of an analytic reflectance function to measured appearance data. This has the benefit of providing significant compression ratios and these analytic models are already fully integrated into modern rendering algorithms. However, this approach can lead to significant approximation errors for many materials and it requires computationally expensive and numerically unstable non-linear optimization.
An alternative approach is to compress these datasets, using algorithms such as Principal Component Analysis, wavelet compression or matrix factorization. Although these techniques provide an accurate and compact representation, they do have several drawbacks. In particular, existing methods do not enable efficient importance sampling for measured materials (and even some complex analytic models) in the context of physically-based rendering systems. Additionally, these representations do not allow editing.
In this thesis, we introduce techniques for acquiring and representing real-world material appearance that address these research challenges. First, we introduce the Inverse Shade Trees (IST) framework. This is a conceptual framework for representing high-dimensional measured appearance data as a tree-structured collection of simpler masks and functions. We use it to provide an intuitive representation of the Spatially-Varying Bidirectional Reflectance Distribution Function (SVBRDF) that is automatically computed from measured data. Like other data-driven techniques, ISTs are more accurate than fitting parametric BRDFs to measured appearance data, but are intuitive enough to support direct editing. We also introduce a factored model of the BRDF optimized to support efficient importance sampling in the context of global illumination rendering. We demonstrate that our technique provides more efficient sampling than previous methods that sample a best-fit parametric model.