I became interested in this conference proceedings because of my interest in the texture recognition problem. I had seen a reference to the Kajiya and Kay paper on texture synthesis. When I called ACM for a reprint of the paper I somehow got trapped into reviewing the entire volume.

The conference included 16 sessions, ranging from the usual “Opening (plenary) Session” through 11 sessions in the technical papers program to 16 discussion topics distributed over 4 panel sessions. Because the contributed papers had to be submitted in advance, they are illustrated in all their full-color splendor on coated paper. As is usual in a conference where the VIPs give the keynote speeches and direct the panels, no texts appear--only some rather unsatisfactory abstracts.

Flipping through the volume, it is easily seen that here are only Sun and similar workstation-class screens; no PC VGA terminals will satisfy these artists. The current state of computer graphics is truly amazing. Let me go first to the paper that got me into this adventure: “Rendering Fur with Three Dimensional Textures,” by J. T. Kajiya and T. L. Kay (pp. 271–280), in the session entitled “3D Textures.” The last page of that paper contains two images of a most realistic teddy bear--it is absolutely indistinguishable from the real thing. After describing a brute force method for generating the illusion of fur and hair by generating thousands of polygons or thousands of little cylinders, the authors say

The brute force method fails because the desired detail should be rendered through textures and lighting models rather than through geometry. What is desired is the painter’s illusion, a suggestion that there is detail in the scene far beyond the resolution of the image. When one examines a painting closely the painter’s illusion falls apart: zooming in on a finely detailed object in a painting reveals only meaningless blotches of color.

Further insight is given by K. Perlin and E. M. Hoffert in their paper, “Hypertexture,” on pages 253–262:

In computer graphics objects are traditionally modeled as sets having infinitesimally thin boundary surfaces. Often a computed or digitized texture or displacement is then mapped onto the surface for enhanced realism. However there are limitations in treating object boundaries merely as surfaces.

Many objects, such as fur or woven materials, have a complex definition which is at best awkward, and at worst impossible, to describe by a surface model. For other objects, such as eroded materials or fluids, a highly complex boundary is actually an artifact of a process that is often more readily described volumetrically. Still other objects, such as flame, clouds, or smoke, don’t actually have a well defined boundary surface at all.

I am impressed by the insistence on the spatial (three-dimensional) nature of the perception of texture in both of these papers, unlike the flatland approach in texture work by the pattern recognition people.

Flipping through the volume, I became more and more amazed. In the session on geometric modeling, Ned Greene (“Voxel Space Automata: Modelling with Stochastic Growth Processes in Voxel Space,” pp. 175–184) describes work with polygons--tens and hundreds of thousands of them--that reproduces vegetation most beautifully; Figure 9A is a bower with a trimmed shrub that looks like a photo of an architect’s model (indeed, the figure title is “Model from ‘Organic Architecture’”).

But all this beauty comes at a price. Greene’s bower took 35 hours on a Sun4. Kajiya and Kay come off a bit better: for 1.3 MPels their program took “only” two hours--using a mere 12 3090s and 4 3081s.

Hence the papers in the “Hardware” session assume great importance. These papers discuss parallel architectures. Unlike earlier generations of massively parallel SIMD Machines (“MPP” and the like), here we find MIMD machines in Michael Potmesnil and Eric M. Hoffert’s paper “The Pixel Machine: A Parallel Image Computer,” pages 69–78, and Henry Fuchs et al.’s “Pixel-Planes 5: A Heterogeneous Multiprocessor Graphics System Using Processor-Enhanced Memories.”

Real time is still elusive, but reality in the images is here.

The other sessions covered the following:

• Natural Models

• Rendering

• Graphics Interfaces

• Visualization

• Ray Tracing

• Radiosity (apparently this term means computing the actual variations in brightness of a scene, taking into account all sources, reflectors, and diffusers)

• Graphics Algorithms

The specialist will need to make an intensive effort to keep up with the field, starting with the annual reviews. The dilettante can be satisfied by looking over such a volume two or three times in a decade.