On Recognizing Faces

An example of the constructed faces
used by HR Wilson in his research.
(b) is derived from (a).

Although humans have no problem identifying them from all orientations, faces look extremely different depending on the angle from which one views them. That is to say that despite our facility in dealing with this challenge, it is an extremely difficult problem to solve with computers. Further, because we have some understanding of the way in which vision is hierarchically organized, progressively building up complex forms starting from small line segments, it points out a formidable computational ability that our brains possess.

Hugh R. Wilson spoke on this issue on Monday (6/16/2008), at the SUNY College of Optometry on 42nd Street. He showed that humans compute facial orientation as suggested by the fact that a subject's sense of this orientation can be adapted by a brief (5 seconds) exposure to a face at a given angle. In other words, before adaptation, humans can fairly accurately identify facial orientations, and this sense is shifted by the adaptation procedure. This basic finding is probably employed to compute what the face would look like from the front, thusly making identification possible.

Of course this psychophysical experiment is only suggestive of such a mechanism, but it is an intriguing possibility.