ORLANDO, Fla. — Foreseeing Hollywood-class computer graphics as the potential killer app that could reenergize the flagging mainstream PC market, Intel Corp. and Microsoft Corp. used the recent Siggraph '98 to disclose separate, ambitious research efforts dedicated to pushing the envelope in 3-D animation and rendering software.

The goal: to make 3-D graphics on systems with Intel CPUs and Windows NT a competitive threat to the Unix workstations that now dominate the high-end market.

For Intel, the Siggraph revelations had additional significance. They marked the official coming out of a newly formed graphics-research group focused on computer animation and advanced modeling. "One of our motivations is to get this advanced technology out into the marketplace," said Demetri Terzopoulos, a member of the group and a professor of computer science at the University of Toronto.

Equally important, the work detailed in a half-dozen papers from the two companies effectively launches the second phase of the Wintel world's "battle for eyeballs," as laid out by Intel chairman Andrew Grove in his landmark Comdex keynote speech of November 1996. Grove stressed his belief that killer applications must emerge if the PC industry is to maintain its momentum and dethrone television as the leisure-time activity of choice for the American consumer.

"That's why this is of importance and interest to Intel," said Terzopoulos. "We are trying to do interesting research and send some of it out into the community, and thereby cause [demand] for high-performance processors."

Though separate projects, all of the work showcased at the conference here-Intel gave two papers, Microsoft's research lab presented four-had at their core a common goal: finding new techniques-usually software algorithms-to reduce the computing power required to render an animated scene and paint it on the display. That's important because, even though processing speed continues to double every 18 months or so, there's still not enough horsepower to handle today's most demanding graphics tasks.

That objective was amply evident in the Intel paper by Terzopoulos and his colleague, Intel senior researcher Radek Grzeszczuk, reporting their development of a neural network implemented in software and running on a garden-variety 300-MHz Pentium II PC.

"This is explicitly for high-performance 3-D animation," said Terzopoulos. "We want to be able to populate realistic virtual environments with hundreds of thousands of physically realistic creatures. If you were to attempt this with physical simulation, it would tax any computer today-even supercomputers. We're coming up with an alternative method for, at minimum, animating vast hordes of background characters."

The network, called "Neuroanimator," is intended to vastly reduce the time it takes to do, for example, an animation of a car in motion or a lunar lander touching down on the moon's surface-two examples shown in a video at Siggraph. "The way it's done now is computationally expensive, so people cannot afford to put too many physically realistic models in their animation," said Grzeszczuk.

"Our approach differs radically from the traditional approach of calculating the equations of motion," said Terzopoulos. "We've replaced them with fast emulators." Indeed, the researchers claimed that Neuroanimator is up to 100 times faster than comparable numerical simulation techniques.

Neuroanimator works by building an approximation of the physical model. "It learns essentially to behave like the model" by observing the model in action,said Terzopoulos. Then the software precisely mimics the physical model, he said, but it can be computationally evaluated "much more efficiently. People still think of neural networks as voodoo, but there's no magic here."

The Intel researchers emphasized that their work is a feasibility study, not a product. However, Neuroanimator is based on public-domain neural-net software called Xerion and runs on commodity Intel CPUs. As such, it could become part of a commercial animation software package.

Making faces
Microsoft also wants to boost the realism of animation while cutting the computational requirements. Two of its four Siggraph papers focused on methods for rendering human facial expressions, such as those used when an actor speaks or visually reacts to another character.

The first, entitled "Making Faces," presented a technique for constructing photorealistic, 3-D animations working off data that's been previously captured and stored using human actors. "An obvious application is the creation of believable virtual characters for movies and television," said the five researchers, led by Brian Guenter of Microsoft Research.

"Another application is the construction of a flexible type of video compression," they Guenter's group said. "Facial expressions can be captured in a studio, delivered via the Internet to a user and then reconstructed in real-time on a user's computer in a virtual 3-D environment. The user can select any arbitrary position for the face and any virtual camera viewpoint, and render the result at any size."

A second Microsoft paper took the technology in another direction, disclosing cutting-edge software that can create natural-looking animation by using 3-D morphing to segue between one facial expression and the next.

Microsoft said the work is "the first step toward building a complete, image-based facial-modeling and animation system."

Evaluating radiance
Another technology, called image-based rendering, was the subject of the final leg of the separate efforts from Intel and Microsoft.

Intel's contribution went a long way toward the goal of cutting the need for CPU horsepower. Intel researcher Baining Guo has developed a so-called "radiance-evaluation algorithm" using progressive ray tracing. It can get away with evaluating as few as 6 percent of the pixels in any given scene to create an image that's difficult to distinguish from the original.

For their part, John Snyder and Jed Lengyel of Microsoft Research have developed an algorithm they said is practical for real-time animation of scenes with hundreds of objects, each of which may contain thousands of polygons. The algorithm implements an efficient method of sorting a set of moving geometric objects into a sequence of layered images. These images, in turn, are then composited-put together in a single frame-to produce the final displayed image.

However, what's most interesting about Lengyel and Snyder's work may not be the algorithm itself so much as what it reveals about Microsoft's future plans for PC graphics. That's because their technique, which hinges on a list of objects stored in memory for future compositing, runs directly counter to today's common practice.

Most graphics software relies on a frame buffer-a single memory space that stores all the pixels that are to be displayed. However, as Jim Blinn, a fellow at Microsoft Research, pointed out in his Siggraph keynote speech last Wednesday, typical applications use multiple windows, which they manipulate and move around the screen a lot. For such apps, the contents of the frame buffer must be recomputed every time an object moves. Blinn dubbed this situation "the tyranny of the frame buffer."

The other way to handle changing scenes is to keep a list of on-screen objects in memory. Objects that exit or enter the scene are subtracted or added to the list. This way, only the changed portion of the screen needs to be recomputed. This is the method Snyder and Lengyel used.

As Blinn noted in his talk, it's also the technique used by Talisman, the 3-D graphics architecture for PCs that Microsoft proposed two years ago at Siggraph '96 in New Orleans. Talisman was intended to make high-end graphics a commodity on the PC based on the use of a media processor and advanced software from Microsoft. Intel also lent its support to Talisman.

Despite the splashy introduction and initial high hopes, Talisman has languished. As the Microsoft paper made clear, however, some concepts from that technology are still in play.