Winter 2011
RESEARCH

Setting the Computer's Sights

...continued from previous page.


How Humans Interpret Faces

Imagine this. You hand your passport over at a border crossing. You have dyed your hair and changed your haircut since you took your passport picture nine years ago. Are you the same person as depicted by the passport picture?

"Our visual systems are very adept at detecting and recognizing faces around us, even at a young age. As humans, we are able to lock onto features which may be statistically very small, yet we perceive them as very strong keys for recognition." Tal says.

How Computers Interpret Faces

Computer vision aims to extract information from digital images and videos about the 3D scene depicted in them. Tal Hassner explains, "we basically want to teach the computer to do everything we do with the visual system."

Can the computer be taught to see images as the human eye does? The complexity and challenge of this conundrum has intrigued the scientific community for years. Indeed, the very notion of perspective has intrigued peoples for nearly two millennia – the Greeks, Romans, Renaissance artists like Leonardo da Vinci, and many others.

Computers attack this problem by assigning two sets of numerical values to the picture that they see. Each time the visual changes – the background illumination, the profile – these numbers change.

Up until about six years ago, algorithms designed for recognizing faces and images had enjoyed quite good results. But, these results were largely based on controlled images (i.e. whether the person being photographed is cooperating and the imaging conditions are either known or at least fixed, such as is typically the case in biometric systems.) The question then arose how to extend this performance so as to enjoy solid results even in cases where the individual is not collaborating and conditions may vary freely. Merely by raising the question, the issue was complicated significantly.

Page: 1  2  3