David D. Hoffman
PERCEPTION: SEEING THE WORLD OR SEEING OUR THEORIES?
Vision seems effortless and direct. We simply open our eyes and observe
the world. With a glance we can perceive the shapes of complex objects, whether
those objects be familiar or whether, as in the figures below, they be less
familiar. But this apparent ease and directness is deceptive. Note, for
instance, that although the lines in the figures below must in fact all lie in
the plane of the page, they appear instead to undulate in three dimensions.
Your visual system has invented the undulations in depth. Note also that
although the two figures are identical (except that one is rotated 180
degrees), their organizations into hills and valleys is nevertheless quite
different. For the figure on the left the dashed contours lie in valleys,
whereas for the figure on the right they lie on hilltops. Not only does your
visual system invent the depth you see, it organizes its invention according to
rules which, evidently, depend on the orientation of the figure.
These
figures are a problem for the "directness" of perception. Direct misperception
seems a contradiction. But are these figures just special cases? Or is
perception more generally, in more ecologically natural situations, also a
process of quick, clever, and perhaps unconscious, invention? In this talk we
explore the structure of current theories in robotic and human vision looking
for answers to these questions. In the process we discover that all theories of
specific perceptual capacities (such as stereovision, edge detection, and the
perception of surface colors) share a common formal structure. Consideration of
this formal structure does indeed provide guidance for the resolution of these
questions. It also allows us to examine with greater precision the problem of
the relation between the observer and the world.