These are technically called Eigenfaces - I first read about them in Scientific American back in the '90s (they were called eigenheads back then) and they looked way freakier. The original research I read basically said there were something like 70 eigenfaces and everyone's face was nothing but a cross of three or less eigenfaces, much like any vector is the sum of two eigenvectors. This is absolutely the way CIA/FBI/NSA facial recognition works. The paper I read was funded by DARPA if I recall.
Here's the article. if someone can get past the paywall, godspeed.A
ccurately re-creating a three-dimensional
face from the subtle shading in a photo-
graph has long challenged computer scientists.
Their algorithms, it now seems, were too gener-
al—aspiring to describe the moon’s surface as
well, or rather as poorly, as the human head. But
by recognizing the fact that head shapes are as-
tonishingly regular, Joseph J. Atick, Paul A. Grif-
fin and A. Norman Redlich of the Rockefeller
University have found a quick means of repro-
ducing the unique contours of a person’s face
from a snapshot.
The discovery may revolutionize the treat-
ment of burn victims. Clear plastic masks, fitted
over a patient’s face to control the formation of
scar tissue, end up determining his or her ap-
pearance. Currently the masks are made by tak-
ing a painful plaster cast of the burnt face. The
Rockefeller technique will instead allow the
masks to be constructed from a photograph tak-
en prior to the burn, by generating the three-di-
mensional face. Scientists at the Computerized
Anthropometric Research and Design (CARD)
Laboratory at the Wright-Patterson Air Force Base
are working to develop such masks.
The insight that led to this breakthrough may
be even more informative. “Any human face is a
combination of a few dozen primary shapes,”
Atick maintains. The researchers analyzed 347
three-dimensional scans of heads of air force pi-
lots—mostly white men—taken at the CARD lab. From 200
of these, they derived an average, adult white male head
shape—dubbed the meanhead (top row, far left )—and a
set of 200 standardized variations from that shape, the
eigenheads (15 of which are shown in consecutive rows).
The latter are so called because they are eigenfunctions,
solutions to a set of linear equations that offer the most
economical way to store information. The eigenheads thus
vastly simplify the derivation of a full face from the shading
in a picture, a problem that would otherwise involve an in-
finite number of variables. Each of the remaining 147 heads
in the database was reproduced to within 1 percent by com-
bining the meanhead with no more than 40 eigenheads.
The eigenheads may be more than a mathematical aid.
The inferior temporal cortex has “face cells,” neurons that
fire selectively when a human visage is presented. Why
certain cells respond to a given face is not known. But
brains have a penchant for eigenfunctions: color, for ex-
ample, is analyzed via the blends of red, green and blue
that form eigencolors. Our brains may also have figured
out that head shapes are best coded as eigenheads. “Each
cell might fire in response to a particular eigenhead,” Atick
suggests—giving humans their incredible capacity to rec-
ognize individual faces. —Madhusree Mukerjee