You literally just set the object on a table and wave your magic wand so it casts its shadow across the object. A few passes gives you a decent picture that, on closer inspection, is as cratered as any comet. But the more passes you make, the smoother the picture gets. And you can change the wand's angle, direction, and speed, or make extra passes over tricky details—as long as both ends of the shadow fall on the desk, the system will work. Scanners need accurate (and expensive) motion control to define the relative positions of the camera and the object, but Bouguet exploits Euclid instead. The lamp, the stick, and the shadow all lie in a plane that intersects the tabletop. Thus the difference between where the shadow lies on the object and where it would have fallen in the background provides the depth.

So the computer scans the top and bottom row of pixels in each frame to nd the shadow's leading edge in the background at that instant. Another part of the system tracks each pixel individually to see when it turns from light to dark, meaning that the shadow has just reached it. The system notes the time, looks up the background shadow points in the corresponding frame, and triangulates where the suddenly overcast pixel is. Standard methods for finding shadows (and other edges) look for abrupt changes between the relative brightness of all pairs of pixels within a set distance of each other, which takes tons of processing time and can be thrown off by surcial color changes or brightness changes, among other things. But here, says Bouguet, "Each pixel raises its hand, saying, I see the edge now! Compute me! And time is insensitive to variations in the scene." (He later learned that Brian Curless and Marc Levoy at Stanford had proved this mathematically two years earlier.)

A line and a point define a plane, so you need to know where the lamp is. Bouguet uses what he calls the Inverse Thales Experiment, explaining "Thales assumed that the light came from a known direction, and wanted to measure the height of a pyramid by comparing its shadow to that of a man of known height; we start with a known height—a pencil—and want to locate the light source.And if we do this several times while moving the pencil around, it gives us several lines that converge back at the lamp."

A newer version doesn't even care where the lamp is. If a shadow falls on two perpendicular, planes—say the table and the wall behind it—the light source can be derived from that information alone. (Two lines may also determine a plane.) You can scan really big objects outdoors, using the sun, as Bouguet demonstrated by scanning Perona's car in front of a handy wall. It doesn' st even matter that the sun moves, because each frame stands on its own. "If youre lazy," says Bouguet, "you could drive a stick in the ground, or even use the shadow of a building, and wait for the shadow to move across the scene."

The method isn't perfect. It can't handle black surfaces, such as Perona's tires, or shiny surfaces, like his windshield, which reflect rather than scatter light—but then, neither will most laser systems. (It does handle nubbly textures much better than the lasers, which require fairly smooth surfaces.) And it only sees what's lit, so areas that are in shadow the whole time don't show up. Nor does the object's back. Thats where active lighting is better, Bouguet admits, because you can see the object from all angles. We could merge several scans from different viewpoints to get a complete 3-D model with no shadow gaps, but there's still significant work to be done in making sure that the errors don't accumulate and globally deform the structure, the way the Beckman Institute hallway became a spiral staircase. But for many home-computer and Web uses, getting 3-D scans for free sure beats buying one of those fancy systems. The process has been patented, and—surprise! a company is interested.

But Perona's vision of machine vision goes beyond computers per se—anything with a chip in it is fair game. He foresees "toys that recognize the child that owns them and are able to play hide-and-seek with her, and washing machines that start when we leave the room and quiet down when we come back so as not to disturb us." He then adds a more serious note. If all cameras become smart, are we on our way to a world where a citizens every move will be tracked automatically, as George Orwell predicted in 1984? The technology to do so will certainly be in place soon, so we as a democratic society had better start thinking about how we plan to regulate what can be done with that information. Being able to interact with a vision-based computer as if it were another human being has a lot of advantages; we just have to make sure that they aren't misused.