Illumination & Textures

17 August 2001
by Forrester Cole

The Illumination & Textures session brought together two papers on techniques for rendering specific objects with two more general papers on improvements to texture mapping methods. The first illumination paper was on the realistic rendering of knitwear, such as wool. The second was a collection of research on methods for modeling a realistic night sky. The first texture mapping paper was a on a general improvement to texturing progressive meshes, where the second was a specific technique for applying texture maps constrained to features of a model.

Photo-Realistic Rendering of Knitwear
Baining Guo showed a new technique for rendering wool and other loosely woven cloth. The work was done by Ying-Qing Xu, Yanyun Chen, Stephen Lin, Hua Zhong, Enhua Wu, Guo, and Heung-Yueng Shum, all of Microsoft Research China.

A central problem with rendering knitwear is that knitwear is made up of many tiny fibers. The small fibers are loosely woven, so that each fiber itself is visible and contributes to the look of the knit. Rather than model each fiber separately, the authors chose to abstract the form of a string of yarn to a two dimensional lumislice. A lumislice represents the light reflected by a slice of the yarn for a given light direction. Computing the lumislice for a given light direction captures the tiny shadowing caused by the individual fibers. To reconstruct the yarn, the lumislice is extruded along the yarn’s axis, rotated to simulate the lie of the yarn. The high level shadowing caused by whole yarn threads is then added into the final image.

The results of the method are convincing. The tiny detail on the knitwear is apparent, and the overall appearance is good, if not quite photo-realistic as advertised.

A Physically-Based Night Sky Model
This paper had only one fewer author than the previous one. Henrik Wann Jensen of Stanford presented the work, which he wrote with Michael M. Stark, Simon Premoze, and Peter Shirley of the University of Utah, and Fredo Durand and Julie Dorsey of MIT. The motivation for their work was to create an easy to use model that captured the beauty and complexity of the night sky. The paper was more a collection of techniques than a unified model, but the results were indeed pretty.

Jensen stressed the importance of making a model correct in its physical units. The amount of light in different parts of the night sky varies over several orders of magnitude. If the scene is too dim, you cannot simply double the amount of light and expect to get good results. The relative magnitudes will not be preserved.

The authors’ model attempts to faithfully recreate the Moon, atmospheric scattering due to dust, individual stars, and the glow of the Milky Way. Besides creating a large scale physical model, the authors also model the effects of human night vision. One dramatic phenomenon is the shifting in the observed spectrum towards blue. This is an empirically verified effect, and artists have used palette shifts to represent night and low light scenes. Without a blue shift, a simply darkened night scene can look like an under exposed photograph.

The final results of the work appeared believable. However, some color quantization problems (perhaps to be expected, given the almost entirely dark blue palette) tarnished the images.

Texture Mapping Progressive Meshes
The third paper took a dramatic turn from the previous subject to a general method for improving texture mapping. A progressive mesh is a mesh representation well suited to compression and level of detail adjustment. Conventional methods of texturing a progressive mesh can introduce excessive stretching artifacts. Pedro Sander of Harvard presented his work on addressing the problem. The paper was co-authored by Steven Gortler of Harvard and John Synder and Hugues Hoppe, both of Microsoft Research.
The authors introduced a metric that measures the stretching caused by a certain texture parameterization. They also presented an algorithm that can be used to create a parameterization that minimizes the stretching artifacts. The results of using the new metric are textured models with texture skin that is more consistently detailed than unadjusted models. The demonstrations that Sander presented were subtle, but showed improved detail in some difficult areas of high curvature.

Constrained Texture Mapping for Polygonal Meshes
Bruno Levy of INRIA Loria finished the session with the second texture mapping paper. His system is designed to allow a user to associate details of a texture map with details of a model. One of Levy’s examples for the concept was a cow’s head textured with the face of a tiger. The basic approach is to warp the texture in a separate step prior to applying the texture to the model. The user sets the parameters for the pre-warp step by placing feature points on the texture and on the model. One audience member complained that the work seemed to duplicate features already available in commercial packages such as Maya. Levy responded that while it is true that his method and previously available commercial methods perform similar tasks, his optimizations make the concept more intuitive and allow a user to perform a texturing task that might take half an hour in only a couple minutes.