THE VISUAL PERCEPTION OF SURFACE
ORIENTATION FROM OPTICAL MOTION
By James J. Todd and Victor J. Perotti
Presented by Samuel Crisanto
What motivates the paper?
Scientists were investigating what information is available
from motion by studying how people interpreted a series
of frames in an animation.
In particular, they were wondering:
● How much information is uniquely specified in the optic
array?
● How much of that information do human perceptual
systems extract?
What information is uniquely
specified in the optic array?
It was shown that:
● You can specify rigid motion from non-rigid motion given
○ 2 views (frames)
● You can mathematically determine an object’s 3D
structure given
○ 3 distinct views of 4 points
Do our perceptual systems
extract this information?
When you show people a multiple frame sequence:
● They cannot differentiate between structures even
though there is sufficient information
● They see a definite 3D shape even if available
information is ambiguous
○ They give a “highly reliable answer…[that exhibits]
systematic biases”.
What do we want to know?
● The paper asks:
○ What aspects of image motion determine how we
perceive the orientation of a flat, rotating surface?
The Claim: Our visual perception of 3-D structure from
motion is based primarily on first-order temporal
derivatives of moving images.
First Order Temporal Derivatives?
VELOCITY is first-order information
•Our perceptual system cares about how quickly points
move
•Our perceptual system does not care about how quickly
points accelerate (this is higher order information)
What information is uniquely
specified in the optic array?
We can model a surface as
______________________________
Tilt is uniquely specified
Tilt =
______________________________
Slant is not uniquely specified
(we need the angular velocity w)
Slant =
What information do our
perceptual systems extract?
The Hypothesis:
● People can accurately judge tilt because it is a ratio of
first-order relationships.
● People base their estimation of slant on the local
deformation of textures.
(Recall that slant =
)
The Experimental Stimulus
People used shuttered glasses to view an LCD screen with
a stereogram of a rocky texture
The texture simulated two
oriented surfaces
The texture had been mapped to a dihedral angle
They deformed the texture
to simulate rotation
● They occluded the edges so observers could not get
information from bounding contours
Observers reported perceived
tilt and slant
● Observers could toggle back and forth between the
deforming texture and a computer model of a dihedral
angle
● They were allowed to adjust the model of the angle until it
matched what they perceived
Experiment 1:
● Models a constant flow field: no higher order
derivatives. Slant is NOT uniquely specified in a
constant flow field.
● 26 different display conditions, with different velocity
values.
● The display conditions were presented five times each
in a random sequence over a period of two
experimental sessions.
Experiment 1: Tilt in a
Constant Flow Field
Observers are almost perfectly accurate at determining tilt
Experiment 1: Slant in a
Constant Flow Field
Perceived slant depends on Vx and Vy, not deformation
A New Hypothesis
People do not judge slant based on deformation. They
instead judge slant by
where tau is tilt and alpha is a free parameter.
IMPLICATION: Since slant is underdetermined, our visual
system uses tilt, something which is uniquely determined.
Support for the Hypothesis
Observer’s perceptions fit the new hypothesis (alpha = .24)
Experiment 2:
Higher Order Relations
● The texture no longer deforms with a constant flow field:
higher order relationships were simulated
● We now have a true simulated slant
● Experimenters used two different angular velocities
● If we use higher order relations, then the different
values of omega will yield different perceived slant
Experiment 2:
Perception of Tilt
People are still very accurate perceivers of tilt.
Experiment 2:
Perception of Slant
People are less accurate at perceiving slant.
Experiment 2:
Testing the old hypothesis
Our perception of slant is not based purely on the
deformation of a texture.
Experiment 2:
Testing the new hypothesis
Using .066 as the free parameter
Conclusion
● The perception of 3-D structure from motion is
based primarily on first-order temporal derivatives
of moving elements within a visual image.
● People accurately estimate tilt, which is uniquely
specified by first order information
● People inaccurately judge slant, even when slant is
uniquely determined by higher-order derivatives.
○ People seem to judge slant as a function of first
order derivatives and tilt.