1
The object in A can be
recognized over:
C – translation, as in
movement of the object or
gaze
D – changes in size
(distance from observer)
E – changes in lighting
(from upper left to upper
right)
F– shifts in picture-plane
orientation
G-shifts in depth-plane
orientation.
H&I – viewpoint shifts
caused by SOGI turning
in depth or observer
moving around the object.
Object Constancy refers to our
ability to recognize different two-
dimensional images as representations
of a particular three-dimensional
object (SOGI, in this case).
• Note that changes in viewpoint might cause parts
of the object to become visible or to occlude other
parts of the object.
– The star-shaped side of SOGI shown in H occludes
more of the figure than in A.
• To a computer, all of the images in figure 1 are
just collections of numbers that specify the
intensity of each small dot, or pixel.
– Teaching a computer to distinguish between the figures
in A and B would be relatively trivial because all that
would be required would be point-by-point
comparisons of pixels, but teaching the computer that C
through I are also the same figure as A would be much
more difficult.
– Pixel-by-pixel comparisons would also find A to be
different from images in C through I.
Template Matching
• Before considering 3-D figures, let us examine
possible representations of two-dimensional
patterns.
•Template theories propose that patterns are not
really analyzed at all—templates are holistic
entities that are compared to input patterns to
determine amount of overlap.
• Template matching works well in pattern
recognition machines that read letters and numbers
in standardized, constrained contexts (scanners that
read your account number off from checks,
machines that read postal zip codes off from
envelopes).
Template Matching Process for Letters: The letter must match the
template exactly as in (a). The template matching procedure can fail
because of (b) change in position, (c) change in size, and (d) change in
orientation.
Template Matching: Strengths
• There is abundant physiological support that
simple features (lines and edges of
particular orientations) are represented in
the nervous system with template-like
receptive fields in the visual cortex.
• They are amazingly reliable. If the to-be-
encoded stimulus is present, it’s template
will become active.
Template Matching: Weaknesses
• The difficulty with template matching as a
model for perception is that contexts are
rarely constrained.
• For instance, slight deviations in shape, size,
orientation, would prevent template
matchers from reading even the limited
number of letters (26) in English
– They are not inherently view invariant. For
every different possible view, there would have
to be a different template (replication). As such,
template representations are uneconomical.
