Acquisition and modeling of material appearance

Abstract

In computer graphics, the realistic rendering of synthetic scenes requires a precise description of surface geometry, lighting, and material appearance. While 3D geometry scanning and modeling have advanced significantly in recent years, measurement and modeling of accurate material appearance have remained critical challenges. Analytical models are the main tools to describe material appearance in most current applications. They provide compact and smooth approximations to real materials but lack the expressiveness to represent complex materials. Data-driven approaches based on exhaustive measurements are fully general but the measurement process is difficult and the storage requirement is very high. In this thesis, we propose the use of hybrid representations that are more compact and easier to acquire than exhaustive measurement, while preserving much generality of a data-driven approach. To represent complex bidirectional reflectance distribution functions (BRDFs), we present a new method to estimate a general microfacet distribution from measured data. We show that this representation is able to reproduce complex materials that are impossible to model with purely analytical models.

(cont.) We also propose a new method that significantly reduces measurement cost and time of the bidirectional texture function (BTF) through a statistical characterization of texture appearance. Our reconstruction method combines naturally aligned images and alignment-insensitive statistics to produce visually plausible results. We demonstrate our acquisition system which is able to capture intricate materials like fabrics in less than ten minutes with commodity equipments. In addition, we present a method to facilitate effective user design in the space of material appearance. We introduce a metric in the space of reflectance which corresponds roughly to perceptual measures. The main idea of our approach is to evaluate reflectance differences in terms of their induced rendered images, instead of the reflectance function itself defined in the angular domains. With rendered images, we show that even a simple computational metric can provide good perceptual spacing and enable intuitive navigation of the reflectance space.