Garner Interference as a Diagnostic for LowLevel Visual Grouping

1
Garner Interference as a Diagnostic for Low-Level
Visual Grouping Emergent FeaturesJames R.
Pomerantz, Rice University, Houston, Texas, USA
Abstract Garner Interference (GI) has been shown
to measure the grouping of line segments into
unitary perceptual configurations. When
participants classify curved line pairs, such as
((, (), )(, and )), according to one of the
lines, RTs rise sharply when the other,
task-irrelevant line varies orthogonally.
Altering the lines positions (and thus their
perceived grouping) changes their GI levels in a
systematic, orderly fashion. We extend these
findings to simpler forms of perceptual
organization with as few as two dots in the
field, and we explore GIs potential to identify
such emergent features as symmetry, collinearity,
parallelism, closure, relatability,
intersections, terminators and edges. Lastly, we
contrast GI with other diagnostics for grouping,
including Stroop-like incongruity-based
interference (as compared with Garner, which is
variation-based) and Configural superiority
effects.

• Theory
• If two elements group perceptually, three
  consequences should ensue
• 1.   Selective attention to just one of the
  elements should not occur, either because it is
  impossible grouping involves a gluing together
  of elements or because it is merely inefficient
  compared with distributing attention across
  elements to detect emergent features such as
  symmetry, closure, or good continuation
  (relatability). In theory, selective attention
  can be measured by Garner Interference or by
  Stroop Interference.
• Conversely, divided attention across the elements
  should be easy.
• Identification or discrimination of an element
  should be facilitated by the presence of a second
  element, even if that new element conveys no
  task-relevant information. This can be measured
  by Configural Superiority Effects.

Parenthesis Variations Figure 4 shows several
variations on the basic parenthesis pairs that we
have tested. These include separating the
parentheses, rotating them, enlarging or
shrinking them, and adding context lines between
or around the elements. Garner Interference has
responded to these variations in systematic ways.
E.g., separating the elements horizontally yields
a monotonic decrease in Garner, as would be
expected via grouping by proximity. Separating
them vertically yields non-monotonic changes,
however, as would be expected via grouping by
good continuation. Joining the elements with a
diagonal line, so as to produce N-like
configurations, weakens but does not eliminate
Garner, but placing the elements in a face-like
context eliminates Garner entirely.

• Driving Questions
• Under what circumstances do two elements in the
  visual field line segments, curves, dots, or
  other contours or blobs group to form a unitary
  configuration? How can this grouping be
  assessed? In this poster, we pursue three
  performance measures of grouping, with emphasis
  on the first
• Garner Interference (GI)
• Stroop Interference (SI)
• Configural Superiority Effects (CSEs)

Method The basic method tests stimulus sets using
the three tasks shown in Figure 1. Stimuli are
presented on a screen, one at a time, and the
task is to press a response key (rather than say
responses silently as in the demonstration) as
quickly as possible according to a rule specified
by the task. Subtracting RTs in the Control Task,
where the irrelevant element stays constant, from
the Filtering Task, where it varies, yields
Garner (variation) Interference. In both the
Control and Filtering Tasks, half of the stimulus
pairs are congruent (the two elements are the
same) whereas half are incongruent. Subtracting
the former from the latter yields Stroop
(Incongruity) Interference.
Demonstrations Experience Garner, Stroop and
Configural Superiority for yourself in the three
tests below in Figure 1. The task is the same
for all three to read down the column of stimuli
as quickly as possible saying A when the
leftmost or only element is ) and saying
B if it is (. Time yourself separately for
each column. Then do these calculations - Garner
Interference Filtering Control- Configural
Superiority Control Singleton- Stroop
Interference cannot be measured numerically
here, but it corresponds to longer RTs for the
incongruent stimuli ) ( and ( ) than for the
congruent stimuli ( ( and ) ).
Stimuli and Results The experiment demonstrated
in Figure 1 has been performed with many stimulus
sets, some of which are shown in Figure 2. Sets
of four stimuli are generated by crossing two
alternative possibilities (lines or shapes,
letters, colors, etc.) for the first element with
two for the second. Figure 2, Panel A shows the
basic parenthesis pairs, along with the level of
Garner Interference they produce, namely 184 msec
per stimulus. This Garner interference is
presumably attributable to the emergent features
of symmetry, closure, and good continuation the
parentheses possess. When the right element is
rotated 90 deg as in Panel B, so as to weaken
grouping by eliminating these emergent features,
Garner drops to an insignificant 36 msec (and has
fallen to zero in other experiments). Panel C
shows circles and squares side by side, yielding
Garner of 86 msec, perhaps attributable to the
emergent feature of sameness vs. difference.
Panel D shows a similar result with side by side
colors, red and green. Panel E shows high levels
of Garner between adjacent letters (although it
does not show that this Garner is highly
asymmetric, running much more strongly from left
to right than vice versa). Panel F shows curves
similar to the basic parenthesis pairs but with
vertical lines added. Garner remains quite
strong at 110 msec, showing that closure is
probably not a key factor in this grouping, at
least compared with symmetry. Panel G shows
another classic set, diagonal and vertical lines
combining to form arrows and triangles and in the
process generating 144 msec of Garner.
Figure 4
Figure 2
Configural Superiority Odd Quadrant
Paradigm Figure 5 illustrates the Configural
Superiority Effect (CSE) superior
discrimination between differing elements when
they are accompanied by identical (thus
non-informative) contexts - in an Odd Quadrant
paradigm. Here the task is to indicate which
quadrant differs from the other three. (In the
examples, it is always the southeast quadrant
that is different). Panel M shows a
well-established CSE arrows are distinguished
from triangles more easily than are the
positively sloped from negatively sloped diagonal
line segments that make arrows and triangles
different. Panel N shows a corresponding
Configural Inferiority Effect, indicating that it
is not just any context that improves performance
but instead a context that adds emergent
features. With arrows and triangles, these
features possibly are closure, terminators, or
intersections. Panel O shows the CSE with
parenthesis pairs. Panels P T show new CSEs
currently being tested, all formed from the same
base discrimination one dot that differs in
relative position from the other three. The
one-dot context added in Panel P yields the
emergent feature of Orientation, whereas that of
Panel Q yields Proximity. Demonstrating CSEs
with these stimuli will establish Orientation and
Proximity as two basic emergent features. It
will also allow direct comparison of the salience
of these emergent features because they arise
from identical base discriminations, namely the
position of a single dot. Panel R provides the
same demonstration of the emergent feature of
Collinearity vs. Non-Collinearity
(triangularity), whereas Panel S demonstrates
whatever features distinguish between a
T-Intersection and a Square/Diamond. Panel T
shows that the CSE with parens pairs can be
established through exactly the same base
discrimination the change of position of a
single dot. Finally, Panel U shows that the
arrow-triangle CSE cannot be achieve through
moving a single dot but requires the movement of
at least two dots. This suggests a potentially
important difference between the emergent
features of parenthesis pairs and those of arrows
and triangles.
Singleton Control FilteringTask
Task Task
New Stimuli and Microgrouping Panels H L of
Figure 2 show new sets of stimuli currently being
tested, along with the levels of Garner
Interference already established in pilot
testing. Although these dot stimuli might appear
too simple to reveal much in the way of grouping,
configuration, or emergent features, recall that
elements such as these underlie much of
Wertheimers early work on grouping in matrices
and in apparent motion. Moreover, stimuli as
basic as these (1) offer the potential of
revealing the most fundamental principles of
visual grouping and (2) may allow assessment of
many putative emergent features on a uniform,
principled basis. The stimuli in Panel H show 120
msec of Garner Interference, showing that
perceivers cannot (or do not) judge the presence
or absence of a dot in one position while
ignoring the presence or absence of a second.
This indicates that the dots are not perceived
independently, i.e., that they form a group.
Panel I shows significant Garner, indicating that
perceivers cannot or do not attend to the
position of one dot while ignoring the position
of another. Panel J shows the same finding with
dot size, and Panel K shows the same about dot
lightness. Panel L reveals that there are limits
to these interference effects perceivers can in
fact attend to the position of one dot while
ignoring the size of another.

• Garner (variation) Interference vs.Stroop
  (incongruity) Interference
• Across a large number of stimulus sets including
  most of those shown here, there is no correlation
  between Garner and Stroop Interference, as
  revealed the scattergram in Figure 3. Our
  working hypothesis is that Garner is a good
  diagnostic for grouping, whereas Stroop is a poor
  one. Here are the reasons
• Garner is found even with stimuli that cannot in
  principle generate Stroop because no pairing are
  either congruent or incongruent
• Many stimuli that appear to group strongly, such
  as the parenthesis pairs, yield large Garner but
  no Stroop.
• In principle, stimulus elements that group
  strongly should not show Stroop, because if two
  elements are fused perceptually into a single
  one, there can be no congruity or incongruity
  (i.e., it takes two or more things to agree or
  conflict).
• Some manipulations that reduce Garner also
  increase Stroop, which would be expected if
  weakening the grouping produced separate elements
  that could then be congruent or incongruent. For
  example, instructing perceivers to focus on just
  one parenthesis of a pair reduces Garner but
  increases Stroop, suggesting that focusing breaks
  these unitary pairs into separate stimuli that
  may then agree or conflict.

Figure 5
Conclusions Garner Interference is a sensitive
and theory-based measure of perceptual grouping,
whereas Stroop Interference apparently does not.
Together with assessments of configural
superiority effects, Garner can help discover and
quantify the emergent features that we believe
underlie perceptual grouping.
James R. Pomerantz, Rice University, Houston, TX
77006, USApomeran_at_rice.edu, telephone
713-348-3419Thanks to Apu Agrawal, Mary
Portillo, Steven Jewell, Hana Khan,and Martha
Jeong for help with the pilot data and faces data.
Figure 3
Figure 1