Efficient Capture of Multispectral Reflectance of Complex Surfaces

Abstract

The synthetic generation of photo-realistic images is a long-standing goal in the area of Computer Graphics. The applications are numerous, including virtual prototyping, advertisement and visual effects in movies. Since the early days of Computer Graphics, a tremendous amount of work has been spent on rendering algorithms and on the representation of scene geometry, light, and surface reflectance. While some of these areas evolved rather quickly, the optical reflectance behavior of surfaces did not receive much attention over a long time.<br /> In the first time, only simple textures and phenomenological models with few parameters based on rather simple assumptions about surface materials were used. Over the time, better physically-based reflectance models were developed which have the potential to describe at least some classes of materials faithfully. The introduction of so-called <em>data-driven</em> representations for reflectance can be considered a major breakthrough. Here, reflectance data is typically acquired from real-world samples and stored in large tables with dense sampling. Data-driven techniques led to a substantial increase in the quality of virtual material appearance especially for complex surfaces exhibiting a lot of structure and spatially varying reflectance behavior. Measured reflectance properties of real-world materials also helped to use reflectance models in a more sensible way, by finding parameters for which the model resembles the real-world sample as good as possible. Afterwards, the parametric model can unveil its strengths, namely memory-efficiency and simple, physically-based editing.<br /> Nowadays, measuring reflectance properties of surfaces has become more commonplace. The use of digital cameras enabled the development of efficient setups to acquire reflectance data with a dense sampling of both angular and spatial domain. In nearly all cases, RGB or similar trichromatic cameras were utilized, because they are cheap and easy to use. Unfortunately, the discretization of the light spectrum using three filters, which might even overlap, leads to inaccurate colors by an effect called <em>metamerism</em> as soon as different light sources and materials are combined during rendering. In principle, this can be easily circumvented using a better sampling of the spectral domain. Unfortunately, the necessary spectral devices are expensive and require a lot of additional effort to provide high quality measurement data. For this reason, densely sampled reflectance measurement of complex surfaces with good spectral resolution still remains an open problem in Computer Graphics.<br /> The methods presented in this thesis are a step towards widely usable spectral reflectance capture. The basic idea is to re-utilize established and matured RGB technology and to add as few spectral measurement data as possible. The dense, spectral reflectance data is reconstructed afterwards using a novel method. This way, the amount of dedicated spectral hardware like filters, cameras, and special light sources and therefore cost and complexity of hardware are reduced, and the speed of spectral measurements is increased.<br /> We divide the work towards that goal into three main steps: The first one is to build and calibrate a measurement setup that can acquire spectral reflectance data in a brute-force manner. This setup allows us to acquire ground-truth data which in turn enables to judge the quality of more sophisticated methods.<br /> The second step is to develop and evaluate a novel reconstruction method for spectral images from sparse spectral and dense RGB data. This method is the foundation of our enhanced spectral reflectance capture. The spectral reconstruction is performed by minimizing an energy function that meters deviation from the measured data as well as compliance with a novel prior.<br /> In the last step we aim at the real usage of our aforementioned method in a combined RGB-spectral measurement setup. We propose a novel solution to the practically relevant problem of obtaining a highly accurate spectral characterization of RGB cameras. We then evaluate a method to rapidly acquire the additionally required spectral data by integration of only one spectral camera into an existing RGB setup without any further modifications. We demonstrate first results for spectral reflectance capture using this setup.