Ph.D Thesis, Department of Computer Science, KU Leuven, Celestijnenlaan 200A, 3001 Heverlee, Belgium, August 2009.
Materials are an essential component of realistic three-dimensional (3D) com- puter graphics. Each material model offers a set of parameters to control every aspect of its appearance. Choosing the right material parameters to obtain a desired look requires some experience. The appearance of a material is also affected by the shape of the object. We present a broad exploratory study of the influence of shape on material perception, using a same/different pair comparison experiment. The influence was found to be significant enough to cause a noticeable difference in the perceived material. While most traditional material editing tools visualize materials on a sphere, analysis of the psychometric function showed that this shape is the least discriminating. Suggested improvements include the use of one of the more discriminating objects examined, or ideally a visualization on the target shape directly. An alternative solution is to compensate for the influence of shape on material perception. The influence was measured using cue combination in a 2-alternative forced choice (2AFC) experiment. The influence was found to cause a constant shift in perception, dependent only on the shape. By applying the opposite shift to the input material, the accurate perception of the material can be restored. As a proof of concept, this correction was built into an existing material editing tool. In the final part of this dissertation we consider how a scene can be observed indirectly through reflections in glossy materials. The concept of visual equivalence was applied to assess the fidelity of dynamic scenes with distorted reflections. Appearance properties such as shape, motion, and lighting were judged using same/different and 2AFC experimental designs. Rotating objects were found to yield equivalence for small distortions. Equivalence for translating objects depends on the plausible motion of highlights and reflections produced by the illumination interpolation method. Throughout this dissertation, we have combined well-thought-out experimental designs and thorough statistical analyses. By using high-quality photorealistic rendering techniques, such as differential rendering, global illumination, and image-based lighting, we ensure that our results are directly generalizable to a wide range of real and virtual applications.
