Modeling And Rendering Fabrics At Micron-Resolution
Fabrics are essential to our everyday lives. Recently advances in physically based rendering techniques and computing power have made it possible to accurately model and reproduce their appearance. This has led to many applications in not only computer graphics but also other fields including industrial design, retail, and entertainment. Traditionally, fabrics are treated as infinitely thin 2D sheets. These surfacebased reflectance models, although being conceptually simple, have insufficient power to describe a fabric's small-scale 3D geometries, such as disorganized layers of fibers in felts. Thus, these models cannot accurately reproduce the fabric's thickness and fuzziness, limiting the level of realism they can offer. In contrast, volumetric models, which has recently been very successful in modeling fabric appearance, explicitly express those fiber-level structures, upon which greatly varying visual effects, from anisotropic highlights to deep textures, emerge automatically. However, volumetric models are generally difficult to build: one would need to specify spatially varying optical properties at high (e.g., micron) resolutions. This dissertation presents a family of algorithms that introduce a brand new way to automatically build and efficiently render volumetric fabric models at micron-resolution. These models capture rich details at fiber-level, yielding highly realistic renderings even under extreme close-ups. Our first contribution tackles the challenge of automated creation of high fidelity fabric models. Our method uses volume imaging techniques to measure a fabric's detailed structural information. Such information is then combined with a photograph to form a complete model through an appearance matching process. The resulting model offers fabric renderings with unprecedented quality. CT scanning fabric samples can be highly time consuming. Our second contribution aims for creating volumetric models for woven fabrics with complex designs under minimal measurement cost. Our approach starts with building a small database of cloth samples with elementary weave patterns. Then, given an input pattern, a volumetric synthesis stage is performed to form the final volume by copying contents from the database while keeping visible artifacts (such as seams) minimized. Fabric models created by this method have been used by textile researchers at Rhode Island School of Design to preview the appearances of their designs. Our third work focuses on efficient rendering of high-resolution fabric volumes. Based on the observation that these volumes contain repeated structures (i.e., multiple instances of the same content in the database), we precompute light transport information for those structures, and a single precomputation can be reused for many designs synthesized from the same database. During the rendering process, the precomputed information is modularly combined through a Monte Carlo Matrix Inversion (MCMI) framework. This method has accelerated the rendering of thick fabrics by an order of magnitude. Finally, we switch gears and introduce high-order similarity relations to computer graphics. This theory originates in applied physics and studies when two sets of material scattering parameters would lead to identical appearance. We introduce a numerical algorithm to utilize this theory in its general high-order form. The practical usefulness of our method is demonstrated using forward and inverse rendering of translucent media. The approaches presented in this dissertation have created fabric renderings with unprecedented fidelity. Remaining challenges include developing more general synthesis technique to support a wider range of fabric structures (such as knitworks) as well as finding more powerful light transport models and inverse rendering algorithms so that the rendered fabrics match photographs under a wide combination of viewing and lighting configurations. We also believe the techniques introduced in this dissertation can provide valuable insights for developing appearance modeling techniques for general material beyond fabrics.
Computer graphics; Material appearance modeling; Physically based rendering
Marschner, Stephen Robert; Van Loan, Charles Francis; Kleinberg, Robert David
Ph.D. of Computer Science
Doctor of Philosophy
dissertation or thesis