Computer technology has grown ever more advanced in recent decades, but we reached an impasse a while back where technology collided with biology in an unexpected way. Trying to create digital versions of human faces usually resulted in something bizarre or downright disturbing. The phenomenon, known as the uncanny valley, is still vexing for the movie and game industries. However, a team led by Abhijeet Ghosh and Paul Debevec of the University of Southern California (USC) has developed a method to make artificial faces even more real, perhaps crossing the uncanny valley. It turns out the answers were only skin-deep.
The human brain is precisely tuned to understand what a face is supposed to look like. These subtle cues are deeply ingrained and when we find them missing, the response is often viscerally negative. It can be as simple as muscles around the eyes contracting oddly, or the way lips part during speech. Science is getting closer to nailing down the mechanical processes, but the USC team is tackling the most challenging aspect — skin.
It turns out modeling the reflection of light on skin is extremely complicated because skin itself is extremely complicated. It’s a patchwork of bumps, pores, blemishes, and tiny wrinkles that creep in as you approach middle age. When these details are missing, digital skin doesn’t look real and we venture into uncanny valley territory no matter how accurate the movements are. The technique being developed at USC more accurately simulates skin in a few ways, the first has to do with the lighting.
Each simulated light source is split into four rays — one that bounces off the epidermis, and three others that penetrate the skin to different depths before being scattered. The result is a more natural sheen with realistic shadows.
To make this technology really work, the team also cranked up the level of detail for CGI skin. Using a special scanner, Ghosh and Debevec took extremely high resolution images of human skin from volunteers’ cheeks, chins, and foreheads. Each pixel in the images contained an area only 10 micrometers across (that’s 0.00001 meters, by the way). At this level of detail, a single skin cell is only three pixels wide on average.
The scans were used to generate incredibly detailed 3D renders of skin. When combined with the new simulated lighting, the results are incredibly impressive. The CGI network of pores and bumps make the faces look almost real as the artificial light plays across them.
There has been intense interest from game developers and Hollywood as this project has proceeded. The CGI blockbuster Avatar used a rudimentary version of the USC technology to make the film’s blue-skinned aliens more realistic. Activision and Nvidia have been collaborating with USC in hopes of developing a software package that can generate photorealistic faces on consumer hardware like game consoles and PCs. The day might be fast approaching that your in-game avatar looks completely real in every way that matters.