How high is visual short-term memory capacity for object layout?

Abstract

Previous research measuring visual short-term memory (VSTM) suggests that the capacity for representing the layout of objects is fairly high. In four experiments, we further explored the capacity of VSTM for layout of objects, using the change detection method. In Experiment 1, participants retained most of the elements in displays of 4 to 8 elements. In Experiments 2 and 3, with up to 20 elements, participants retained many of them, reaching a capacity of 13.4 stimulus elements. In Experiment 4, participants retained much of a complex naturalistic scene. In most cases, increasing display size caused only modest reductions in performance, consistent with the idea of configural, variable-resolution grouping. The results indicate that participants can retain a substantial amount of scene layout information (objects and locations) in short-term memory. We propose that this is a case of remote visual understanding, where observers' ability to integrate information from a scene is paramount.

Full text

The relation between the mind and the world is of cen-
tral interest in psychology, and scene perception is a multi-
faceted example of this relation. A critical characteristic
of scene perception is the amount of information that
can be extracted from a scene and retained in immediate
memory. Currently, considerable debate centers on this
quantity.
At one end of the debate is the hypothesis that very
little is held in memory beyond an attended object (see,
e.g., O’Regan, 1992; Wolfe, Klempen, & Dahlen, 2000).
In this hypothesis, the intuitive experience of perceiving a
rich world is explained by the richness of the information
available from the world via eye movements (see O’Regan,
1992). The hypothesis of minimal representation receives
support from difficulties of change detection documented
in many studies (e.g., Grimes, 1996; Simons, 1996).
At the other end of the debate is the hypothesis that
rich representations of scenes result from combined con-
tributions of short-term and long-term memory (e.g.,
Hollingworth, 2004, 2005, 2007; Irwin & Zelinsky, 2002;
Melcher, 2006). Long-term memory contributions can be
obviated by experimental design, limiting the represented
information to short-term memory. Visual short-term
memory (VSTM) for objects has been studied carefully,
and many researchers have concluded that it has a sharp
limit, between two and four objects (e.g., Hollingworth,
2006; Vogel, Woodman, & Luck, 2001). However, almost
all of the studies supporting a sharp limit in VSTM ca-
pacity have measured memory for properties of objects.
Object properties constitute only one aspect of scene
memory. A second fundamental aspect is memory for the
layout of objects and other structures in a scene. The rep-
resentation of spatial information is a fundamental com-
ponent of short-term memory models (e.g., Baddeley &
Hitch, 1974; Logie, 1995).
A number of studies suggest that fairly complex lay-
outs of objects can be retained over short-term intervals.
In seminal studies of very short-term visual memory,
Phillips (1974) also included longer short-term intervals.
At the short-term intervals, he found accurate detection
of changes of the position of a single element within
25- element arrays (e.g., above 85% in several conditions).
Simons (1996) contrasted VSTM for object properties and
locations and found location memory, but not object prop-
erty memory, to be near the ceiling (see also Aginksy &
Tarr, 2000; Alvarez & Oliva, 2007; Jiang, Olson, & Chun,
2000). Franconeri, Alvarez, and Enns (2007) recently
concluded that as many as seven locations can be held
in short-term memory. Rensink (2000b) obtained capac-
ity estimates of at least nine for contrast signs of objects.
Brockmole, Wang, and Irwin (2002) presented dot lay-
outs for an integration task and estimated that the number
of dots held in VSTM was about 10. A separate line of
evidence for complex representations of layout is spatial
priming with scenes; Sanocki and colleagues (Sanocki,
2003; Sanocki, Michelet, Sellers, & Reynolds, 2006;
Sanocki & Sulman, 2009) found evidence that broad-
scale representations of layout information are activated
in memory by a prime scene. In sum, multiple sources of
evidence suggest there may be a fairly high capacity for
layout information in VSTM.
However, the relative advantage for layout over prop-
erty memory can be explained by assuming that grouping
processes are involved in encoding and that grouping is
more effective for layout than for object properties. The
idea that nearby objects can be grouped into larger struc-
1097 © 2010 The Psychonomic Society, Inc.
How high is visual short-term memory capacity
for object layout?
Th o m a s sa n o c k i , Er i c sE l l E r s , JE f f mi T T E l s T a d T , a n d no a h su l m a n
University of South Florida, Tampa, Florida
Previous research measuring visual short-term memory (VSTM) suggests that the capacity for representing
the layout of objects is fairly high. In four experiments, we further explored the capacity of VSTM for layout
of objects, using the change detection method. In Experiment 1, participants retained most of the elements in
displays of 4 to 8 elements. In Experiments 2 and 3, with up to 20 elements, participants retained many of them,
reaching a capacity of 13.4 stimulus elements. In Experiment 4, participants retained much of a complex natural-
istic scene. In most cases, increasing display size caused only modest reductions in performance, consistent with
the idea of configural, variable-resolution grouping. The results indicate that participants can retain a substantial
amount of scene layout information (objects and locations) in short-term memory. We propose that this is a case
of remote visual understanding, where observers’ ability to integrate information from a scene is paramount.
Attention, Perception, & Psychophysics
2010, 72 (4), 1097-1109
doi:10.3758/APP.72.4.1097
T. Sanocki, sanocki@usf.edu

1098 Sa n o c k i , Se l l e r S , Mi t t e l S ta d t , a n d Su l M a n
Fourth, typical layout confounds object identity and
location. Objects help represent layout by serving as
landmarks, and location and scale within scenes can help
define objects (see, e.g., Palmer’s [1975] fruit face; Tor-
ralba, 2003). Therefore, identity may contribute to layout
memory, and layout may contribute to identity memory.
In summary, we argue that memory for the layout of ob-
jects in scenes can be fairly high in capacity and involves
four properties: broadness in spatial extent, grouping of
scenic configurations, variation in resolution, and integra-
tion of identity and location information.
Measuring Capacity
The present experiments utilize two of the primary
methods for inferring VSTM capacity. The first is to mea-
sure the effects of display size. In most cases of object-
property memory, there is a sharp drop-off in accuracy
as display size is increased beyond 3 or 4 items, imply-
ing a limit of about 3 items, beyond which objects are
lost from memory (see, e.g., Eng, Chen, & Jiang, 2005;
Hollingworth, 2006; Vogel et al., 2001; for a review, see
Luck, 2008). However, sharp drop-offs in accuracy are
less likely with a memory that is configural and varies in
resolution. In layout memory, capacity limits may cause
gradual changes in fidelity between differing display sizes
and differing situations, rather than dramatic drop-offs.
The second method is a more formal index—namely,
a simple high-threshold model first applied to VSTM by
Pashler (1988). In this model, items are either present in
memory (which supports change detection when one of
those items changes) or missing from memory. When none
of the items present in memory change, a change guess is
issued at the false alarm rate. When applied to data on
object property memory, the model estimates a memory
capacity k of about three objects, which coincides nicely
with the drop-offs in the display size functions (see, e.g.,
Vogel et al., 2001).
Grouping distorts the meaning of k. As noted, a scene of
n items could be encoded in memory as only a few (, n)
larger grouped structures and, perhaps, as few as one large
structure. The Pashler k would not reflect this, however, be-
cause it counts stimulus items (the units that change in the
experiment). It may be most accurate to think of k in terms
of stimulus information that is held in memory, rather than
as a count of representational units. With these caveats in
mind, we will argue that Pashler’s k, together with percent-
ages correct and examples of the stimulus displays, pro-
vides an intuitively useful and quantitative characterization
of the amount of scenic information held in VSTM.
The Present Experiments
The present experiments were designed to examine the
capacity of short-term memory for layout across a sample
of different stimulus situations. We used stimulus displays
that were broad (.20º visual angle), sometimes dense with
considerable detail (Experiment 4), and that confounded
identity and location (Experiments 1, 2, and 4).
The observers’ task was to detect changes in the loca-
tion of one display element. The element was present in
the first display but moved to a new location in the second
tures is well established in visual memory (see, e.g., Enns,
1987; Hollingworth, Hyun, & Zhang, 2005; Irwin, 1991;
Jiang et al., 2000; Sebrechts & Garner, 1981). Thus, in
layout memory, several proximal objects could be encoded
as a single complex object, reducing the total number of
objects represented in memory (Miller, 1956). For exam-
ple, a complex display of 12 objects might be encoded as
3 four-part objects.
Because of grouping explanations, there is ambiguity in
determining the quantity of objects represented in VSTM.
Nevertheless, we can begin to characterize VSTM capac-
ity. Intuitively, if 12 objects and their layout were held in
VSTM, it would be a fairly complex scene and a substan-
tial portion of a typical scene. This is an answer to the
question, “How much information from a scene can be
held in VSTM?”
The purpose of the present experiments was to explore
and characterize the possible upper limits of VSTM ca-
pacity for object layout. We begin with some ideas about
layout memory and the measurement of object capacity.
Then we explore capacity with three types of stimulus
displays.
Typical Layout
Layout memory is likely to have evolved for layouts
typical in our world. We propose that typical layout has
four main properties that have implications for memory.
The f irst is that typical layout is broad in spatial extent.
Layouts include large-scale structures, such as the horizon
line, buildings, or mountains. The world itself is expan-
sive, and important duties, such as navigating new terri-
tory or providing security for a community, require the
monitoring of areas that extend beyond the foveal field
or even the full field of view. Studies of boundary exten-
sion suggest that memory for a scene layout has the design
feature of extending beyond the immediate visual field
(Intraub, 1997). Broad expanses of information must be
represented in parallel to be optimally useful.
Second, layout often consists of dense configurations of
objects, occluded objects, and various background struc-
tures. To be effective, layout memory must encode dense
configurations efficiently. As noted, there is evidence of
grouping in VSTM, and grouping is a likely mechanism
for the efficient encoding of dense conf igurations (e.g.,
Miller, 1956). For example, the foreground of a home site
might include a yard, a sidewalk, and bushes, all of which
could be grouped into a “front yard” configuration.
Third, because aspects of layout vary in scale and sa-
lience, layout memory is likely to vary in resolution. If a
scene layout is too complex to be fully represented, infor-
mation should not be lost in discrete units, such as objects
or regions, because of the importance of extent; instead,
the broad, parallel nature of the memory could be main-
tained with an overall loss of the fidelity of information.
Usually, the loss of information would begin with more
minor (f iner) details. There is evidence that attentional
and memory resolution varies (see, e.g., Alvarez & Ca-
vanagh, 2004; Awh, Barton, & Vogel, 2007; Franconeri
et al., 2007; Tsal, Meiran, & Lamy, 1995), although not in
all situations (e.g., Zhang & Luck, 2008).

Sh o r t -te r M Me M o r y f o r la y o u t 1099
EXPERIMENT 1
Because previous research on object property memory
found large drop-offs in accuracy beginning at about four
items, we first examined display sizes of four to eight
items. Our manipulation of an organizational factor was
motivated by recent work in perception that has implicated
surfaces—and ground surfaces, in particular—as struc-
tures for integrating object and scene information (e.g.,
He, Wu, Ooi, Yarbrough, & Wu, 2004; Ni, Braunstein, &
Andersen, 2005; Vecera & Palmer, 2006; see also Gibson,
1979). Stimulus elements were selected from an organized
pair of surfaces (ground and sky, Figure 1A) or an unorga-
nized pair of surfaces (random arrangement of the same el-
ements, Figure 1B). The more organized arrangement cor-
responded with regular texture gradients, with rectangles
receding in depth as a ground surface and cloud-like circles
receding in depth as a sky gradient. We expected memory
accuracy and capacity to be higher for displays generated
display. This represents a typical change in layout within
the world, in which an object moves from one location
to another (see, e.g, Irwin, 1991; Phillips, 1974). Within
each experiment, the displays were familiar to observers
because the same set of elements was used repeatedly;
however, the layout of elements changed from trial to trial,
requiring short-term memory for recognizing differences
in the new configuration.
Because grouping is likely to be an important aspect
of the encoding process, we examined several factors that
may influence the organization of the displays. In Experi-
ments 1 and 2, we examined the effects of a potentially
important structure in real-world depth perception, that
being texture gradients (Gibson, 1979). In Experiment 3,
we examined the importance of organization into cardi-
nal rows and columns. In Experiment 4, we examined an
image property thought to be important for segmenting
complex displays—namely, color (see, e.g., Wurm, Legge,
Isenberg, & Luebker, 1993).
50
60
70
80
90
100
468
AB
CD
E
Percentage Correct
Display Size
Regular same
Regular different
Irregular same
Irregular different
Figure 1. Stimuli ( panels A–D) and results (panel E) for Experiment 1. Panels A and B
show the prototype stimulus patterns. Panels C and D show stimulus examples with a
display size of eight. Panel E shows the mean percentage correct in each condition.

1100 Sa n o c k i , Se l l e r S , Mi t t e l S ta d t , a n d Su l M a n
periment provides evidence about the capacity of VSTM:
Memory handled the four to eight items in the displays of
this experiment, with little evidence of limits. Thus, the
information in the displays did not exceed VSTM limits.
Examination of example stimuli (Figures 1C and 1D) in-
dicated that a moderately high amount of scenic informa-
tion was held in VSTM.
An ANOVA was conducted on the factors in the figure.
The main effect of display size was reliable [F(2,10) 5
5.50, p 5 .02, h2
p 5 .52]. The drop in accuracy from four to
eight items was 5.2%, and the drop from four to six items
was 5.6%. Thus, the loss in performance with increasing
display size was modest. Performance at each display size
was also measured in terms of sensitivity: d′ values for
four, six, and eight items were 4.40, 3.52, and 3.33, re-
spectively (Table 1). A′ values are also shown in Table 1.
Pashler’s k provides an estimate of VSTM capacity in
terms of number of stimulus elements. The most valid
estimates are calculated from the data for the largest dis-
play sizes, because smaller display sizes may not reach
capacity. The overall estimated capacity was 7.0 items.
Estimates for individual observers ranged from 6.0 to 7.4
items. Note that the estimates were limited by the number
of items presented, and there was a near ceiling effect.
Had capacity extended beyond eight items, the present
experiment would not have measured it.
The effects of the regularity of the texture gradi-
ents tended to be contrary to our hypothesis: Accuracy
was marginally lower with the regular texture gradients
(92.8%) than with the irregular stimuli (94.6%) [standard
error of the difference (SED) 5 0.7%; for the main ef-
fect of regularity, F(1,5) 5 6.17, p 5 .06]. Inspection of
Figure 1E reveals that, for regular different arrays, there
was a decrement in accuracy at the two larger display
sizes. The increased uniformity in structure and texture
in regular arrays may have biased observers toward same
responses, thereby increasing errors to different displays
by making them less likely than with irregular arrays. This
bias also increased same responses to regular same arrays.
The trade-offs due to bias were apparent in interaction ef-
fects: There were interactions of display size and response
[F(2,10) 5 22.88, p , .001], stimulus regularity and re-
sponse [F(1,5) 5 13.06, p 5 .02], and all three factors
[F(2,10) 5 5.62, p 5 .02]. Also, there was a main effect
of response [F(1,5) 5 9.79, p 5 .03].
Discussion
The present experiment, with near-ceiling performance,
provides some information about the capacity of VSTM.
The stimulus displays used in this experiment (see, e.g.,
Figures 1C and 1D) were within the capacity of VSTM.
The Pashler k was 7.0 of the stimulus elements. The el-
ements are likely to be encoded into fewer than seven
groups. Further experiments will be necessary to deter-
mine the upper limit of stimulus capacity.
The lack of any positive effect of texture gradient regu-
larity was rather surprising. However, there may not have
been enough elements in the stimulus displays to portray
gradients effectively; the displays contained an average of
only three elements each in the ground and sky. Experi-
from these regular gradients than for displays generated
from the same elements arranged in an irregular manner.
During the experiment, the stimuli were blocked by regular
versus irregular display type, and observers were shown
the complete array before the start of each block.
Our complex scenes did not lend themselves to verbal
labeling, and we did not use an auditory secondary task to
discourage verbalization in the experiments. There is no
evidence that verbal labels influence standard VSTM ex-
periments. In studies of object memory, Vogel et al. (2001)
and Liu and Jiang (2005) directly compared short-term
memory performance with and without verbalization and
found no differences in memory performance between
conditions.
Method
Participants
. Six students (3 female) from introductory psychol-
ogy courses at the University of South Florida participated in ex-
change for course credit. A small sample was used in this experiment
because of the near-ceiling effect obtained.
Procedure and Design
. On each trial, an initial stimulus (mem-
ory array) of four, six, or eight elements was presented for 200 msec,
followed by an 800-msec blank interval, and then a second ar ray
(test array) remained on-screen until the observer responded same or
different by pressing a keyboard key. The test array was identical to
the memory array on half of the trials and changed on the remaining
trials (one element offset and a new element onset).
The stimulus arrays were generated by randomly selecting ele-
ments from complete arrays of 16 elements (Figures 1A and 1B).
The complete arrays subtended approximately 21º 3 13º (horizon-
tal 3 vertical visual angle). We generated 120 pairs of arrays—60
of each type (regular and irregular)—with the members of a pair dif-
fering in the position of 1 element. The change in position occurred
within either the upper or the lower half of the display (i.e., the ele-
ment did not move from top to bottom). Changes were as equally
likely to be in the upper region as in the lower region. Examples of
a regular and an irregular array are shown in Figures 1C and 1D,
respectively. Each array pair was used twice during the experiment:
once on a different trial and once on a same trial, in which the first
array of the pair was presented twice. Auditory feedback about ac-
curacy was given after each response.
There were 240 test trials divided into blocks of 20 with array type
(regular, irregular) alternating between blocks in a counterbalanced
order. Testing was preceded by two 10-trial practice blocks with a
random sample of the test stimuli.
Results
Accuracy was near the ceiling, averaging 93.7% cor-
rect. Figure 1E shows percentage correct as a function of
display size, response, and texture regularity. Accuracy re-
mained high across display size, ranging from 97.3% with
display sizes of four, to 91.7% and 92.1% at the larger dis-
play sizes, respectively. The high performance in this ex-
Table 1
Overall Percentages Correct (PC) and Estimates of Sensitivity
(d′, A′), Bias (c, b), and Capacity (k), Averaged Over the
Regularity Condition for Each Display Size in Experiment 1
Capacity
Display Sensitivity Bias Pashler’s Pashler Pashler
Size PC d′ A′c b k Empty Both
4 97.29 4.40 .99 2.03 1.00 3.90 11.69 15.59
6 91.67 3.52 .95 0 1.34 5.19 8.63 13.81
8 92.08 3.33 .96 .07 1.29 7.02 7.02 14.04

Sh o r t -te r M Me M o r y f o r la y o u t 1101
conditions. We extended the Pashler formula by using all
16 locations, and the capacity estimates ranged from 13.8
to 15.6 locations in the three different display-size condi-
tions (“Pashler both” in Table 1). Thus, under alternative
assumptions, capacity remained fairly high.
EXPERIMENT 2
The purpose of Experiment 2 was to further examine
capacity and the influence of texture gradients with dis-
plays that were larger in number of elements and denser.
The complete array had 24 items in total, and stimulus
arrays ranged from 8 to 16 items. Complete arrays and
example stimuli are shown in Figure 2.
Method
The method was similar to that in Experiment 1 and involved the
following changes. Display size was 8, 12, or 16 items. The 120
stimulus pairs were generated from the complete arrays shown in
Figures 2A and 2B. The interstimulus interval was increased to 1 sec
in this experiment, although the duration of the first array remained
ment 2 was designed to provide a more powerful test of the
effects of texture gradients.
One last point is that there are alternative ways to apply
the Pashler model to location memory. Observers might
have remembered empty locations instead of f illed loca-
tions or a combination of empty and filled locations. We
modified the model to examine these cases and found gen-
erally similar estimates of capacity. If only empty regions
were remembered, the number of locations would be the
complement of filled display sizes, ranging from 12 empty
locations at a display size of four objects, to 8 empty loca-
tions at eight objects. The fact that memory performance
is somewhat higher in the largest empty condition of 12
locations is evidence against the idea that the observers
used empty locations. Nevertheless, the Pashler formula
can be reapplied to empty display sizes, and the capacity
estimate at the most valid size of 12 empty locations is
11.7 locations (“Pashler empty” in Table 1). If both empty
and filled locations are used by observers, then the display
size is the sum of empty and filled locations, or 16 in all
50
60
70
80
90
100
81216
AB
CD
E
Percentage Correct
Display Size
Regular same
Regular different
Irregular same
Irregular different
Figure 2. Stimuli (panels A–D) and results ( panel E) for Experiment 2. Panels A
and B show the prototype stimulus patterns. Panels C and D show stimulus exam-
ples with a display size of 16. Panel E shows the mean percentage correct in each
condition.

1102 Sa n o c k i , Se l l e r S , Mi t t e l S ta d t , a n d Su l M a n
EXPERIMENT 3
The purpose of Experiment 3 was to further explore ca-
pacity and grouping, with an emphasis on systematic 2-D
displays. We began with a simple grid of identical circu-
lar elements organized as cardinal rows and columns (see
Figure 3A). Much previous work on stimulus organization
has assumed such organization. Both highly organized
and less organized arrays have been selected from such
matrices, and the more organized arrays have produced
better memory performance (Hollingworth et al., 2005;
Irwin, 1991).
We asked a new question related to organization: What
happens when organization into cardinal rows and col-
umns is disrupted? Regular displays were generated from
the 5 3 5 matrix of dots with constantly spaced rows and
columns and a rectangular frame (Figure 3A) and con-
tained 12 to 20 items. In this condition, there should be a
relative abundance of organizational relations between el-
ements, including colinearity, various types of symmetry,
and relations with the axes of the frame and the environ-
ment. The encoding of these relations should contribute
to VSTM performance. We compared performance for
regular arrays with that for arrays in which cardinal orga-
nization was disrupted. We created a prototype array with
distorted interelement spacing and alignment and rotated
the entire array to oblique axes (Figure 3B). In addition,
the surrounding frame was tilted in the opposite direc-
tion. Oblique orientations are not encoded as strongly as
cardinal orientations, which are more predominant in the
environment (see, e.g., Lasaga & Garner, 1983; White,
Coppola, & Fitzpatrick, 2001). Consequently, the irregu-
lar arrays should contain fewer regular relations and fewer
horizontal and vertical relations that support efficient
memory encoding.
As noted, the present elements were simpler than those
in the first two experiments, and there were no differences
in the identities of the elements. Pilot work suggested that
these stimuli would produce higher capacity than in Ex-
periment 2, so we increased the display sizes somewhat.
Method
Display size was 12, 16, or 20 items. The 120 stimulus pairs were
generated from the two 25-element prototypes shown in Figures 3A
and 3B. For each stimulus pair, element positions were randomly
selected from the relevant complete array. Across pairs, the changing
element stayed within the upper or lower portion of the array half
of the time. The irregular arrays were created by distorting positions
in the regular arrays. The irregular arrays fit within the same square
200 msec. A larger sample of 33 students participated (21 female).
The data for 4 additional participants were not included in the analy-
sis because those participants failed to follow the instructions.
Results
Accuracy levels were intermediate and ranged from
84.7% at display size 8 to 69.9% at display size 16. In the
ANOVA, the main effect of display size was reliable and
moderately large [F(2,64) 5 55.50, p , .001, h2
p 5 .63].
The drop in accuracy from 8 to 16 items was 14.8%. Per-
formance at each display size was also measured in terms
of sensitivity: d′ values for 8, 12, and 16 items were 2.3,
1.7, and 1.3, respectively (averages from Table 2).
The Pashler’s k estimate was a capacity of 8.0 stimulus
items. On a more intuitive level, using the largest display
size as an example (see, e.g., Figures 2C and 2D), observ-
ers were able to hold many of the items in a 16-element
display in VSTM, achieving a level of change detection of
70% for these complex displays.
In this experiment, there was a small positive effect of
regularity of texture gradient. Accuracy was 2.1% higher
with the regular-texture gradients (78.0%) than with the
irregular stimuli (75.9%) [SED 5 0.9%; for main effect of
regularity, F(1,32) 5 5.19, p 5 .03, h2
p 5 .14]. Regularity
of texture interacted with display size [F(2,64) 5 4.01,
p 5 .02, h2
p 5 .11]; the advantage of regularity was reli-
able at the middle display size of 12 [t(32) 5 3.28, p ,
.01], but not at the other display sizes ( ps . .20). The only
other reliable effect in the analysis was the interaction of
display size and response [F(2,64) 5 21.07, p , .001].
Discussion
This experiment confirms that the capacity of VSTM
for texture elements is about 8 stimulus items. As display
size increases beyond 8, the memory representation ap-
pears to lose resolution, lowering overall accuracy to 70%
with 16 elements. The capacity estimates remain moder-
ately high with alternative assumptions (Table 2).
With the denser displays, there was a small positive
influence of systematic gradients. However, the inter-
pretation of this result may be somewhat complicated.
Although gradient structure itself could increase perfor-
mance, note that the regular gradients have small elements
near fixation. This means that element size decreases with
cortical resolution, with the regular gradients. These fac-
tors should be disentangled in future research. In Experi-
ment 3, we continued our sampling of different types of
stimulus displays.
Table 2
Overall Percentages Correct (PC) and Estimates of Sensitivity (d′, A′), Bias (c, b), and
Capacity (k), for Each Display Size in Experiment 2
Sensitivity Bias Capacity
Display PC d′A′c b Pashler’s k Pashler Empty Pashler Both
Size Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg.
8 84.39 85.08 2.32 2.38 .89 .90 2.24 2.02 1.29 1.10 6.11 6.49 12.21 12.98 18.32 19.48
12 79.02 73.56 1.94 1.43 .85 .80 2.29 2.24 1.40 1.37 8.12 6.94 8.12*6.94 16.25 13.87
16 70.68 69.17 1.30 1.22 .78 .76 2.35 2.35 1.51 1.54 8.05 7.58 4.03 3.79 12.08 11.37
Note—Reg., regular; Irreg., irregular. *This value is identical to that for Pashler’s k because these two conditions are complements at display
size 12 (when half of the elements are present).

Sh o r t -te r M Me M o r y f o r la y o u t 1103
rays [SED 5 1.6%; F(1,15) 5 20.91, p , .001, h2
p 5 .58].
The interaction of display size and regularity was mar-
ginal [F(2,30) 5 2.89, p 5 .07]. There was also a main
effect of response [F(1,15) 5 23.04, p , .001].
Pashler ks were 13.4 items for regular displays and 8.7
for irregular displays (Table 3). With regular displays, ob-
servers held many of the elements in VSTM; for example,
16 element displays (e.g., Figure 3C) resulted in an 80%
level of accuracy. Note that the k of 13.4 was obtained
with items that were identical: There was no confound of
identity and location. Also, there was no appearance of
depth in these displays. It appears that a higher number of
simple flat elements can be held in VSTM than can more
varied elements that convey some depth (Experiments 1
and 2).
area as the regular arrays, but because of the irregular orientations,
the interelement distances (including distances between changing
elements) were slightly less than with the regular arrays. As pre-
viously, stimulus regularity was manipulated between blocks, and
observers were shown the appropriate complete array before each
block. Timing parameters were the same as in Experiment 2. A new
sample of 16 students participated (11 female).
Results and Discussion
Performance decreased modestly with display size from
77.7% at 12 items to 73.4% at 20 items, a total drop of
4.3% over an 8-item difference in display size (see Fig-
ure 3E and Table 3). The main effect of display size was
reliable [F(2,15) 5 4.01, p 5 .03, h2
p 5 .21]. In addition,
there was a benefit of stimulus regularity: Accuracy with
regular displays was 6.9% higher than with irregular ar-
50
60
70
80
90
100
12 16 20
AB
CD
E
Percentage Correct
Display Size
Regular same
Regular different
Irregular same
Irregular different
Figure 3. Stimuli (panels A–D) and results ( panel E) for Experiment 3. Panels A
and B show the prototype stimulus patterns. Panels C and D show stimulus exam-
ples with a display size of 16. Panel E shows the mean percentage correct in each
condition.

1104 Sa n o c k i , Se l l e r S , Mi t t e l S ta d t , a n d Su l M a n
objects. In the entire 12-region scene, there was a total of
26 to 40 objects, depending on the criteria for objecthood.
The objects included portions of the railing, a shed, and
a wall. Pictures were taken with the regions filled (all of
the region’s objects present) and empty (all of the region’s
objects absent), and the regions were recombined in Pho-
toshop to create randomly varying layouts. Display size
was varied by using 4, 6, or 8 filled regions. With the aver-
age of 2.17 objects per region (by the most conservative
count), display sizes corresponded to an average of 8.7,
13.0, and 17.4 objects, respectively.
To further examine the nature of the memory repre-
sentation, we restricted changes to occur within 1 of 3
superregions on each trial. As shown in Figure 4D, there
was an outer superregion, in which changes altered the
scene envelope. In addition, there were two superregions
within the house: an inner-front and an inner-back region.
On 33% of the trials, for example, the change was the
deletion of objects in one inner-back region and the ap-
pearance of objects in another inner-back region. If the
memory representation encodes the entire configuration,
changes should be detected within each superregion. If
the memory representation loses discrete objects or re-
gions as capacity is reached, certain superregions (e.g.,
inner-back regions) may suffer most. Alternatively, if the
memory representation is defined mainly by its outer en-
velope, outer changes should be detected much better than
are inner changes.
To explore another potential organizational factor, we
examined the contribution of an image property—the
presence of color—to capacity. The dollhouse and objects
contained a variety of bright colors (e.g, bright pink, yel-
low, blue, and white). Color is believed to be useful for
segmenting complex scenes (see, e.g., Wurm et al., 1993)
and may provide an additional feature that can be used
to encode complex information in short-term memory.
We compared performance on blocks with color images
versus that on blocks with grayscale versions of the same
scenes.
Method
For the color condition, 54 pairs of stimulus layouts were deter-
mined by a random algorithm for choosing regions. The stimuli were
then created by cutting and pasting from a set of highly colorful
digital pictures that together showed each state of each region. Filled
regions were randomly chosen with the constraint that changes (the
deleted and new region) always occurred within one of the three
superregions, each for one third of the trials (see Figure 4D). Fifty-
As can be seen in Table 3, d′ values ranged from 1.8
to 2.2 in the regular conditions and from 1.1 to 1.7 in the
irregular conditions. The difference in sensitivity between
conditions was highly reliable [t(15) 5 4.95, p , .001].
Changes in bias in the main conditions were fairly small
(Table 3). The sensitivity effects indicate that organization
in regular horizontal and vertical matrices contributes to
memory capacity. This result provides new evidence that
organizational factors contribute to VSTM. We suggest
that the greater symmetry and more regular relations of
regular displays lead to more efficient memory coding be-
cause they support the creation of well-organized groups.
The VSTM encoding system may work at least as well
with 2-D relations as with apparent depth relations.
A variety of types of groups may have been encoded
in the regular conditions. First, the outer envelope of the
display is a potentially salient feature, and grouping could
involve the overall shape of this envelope or its compo-
nents (the four sides). Changes in elements could alter the
envelope or open or close holes along the sides. Also, ele-
ments may form salient parts, such as a rectangle or line
(see Figure 3C), and changes in these parts may be strong
signals for change. Systematic study of types of grouping
may be fruitful.
EXPERIMENT 4
Experiments 1, 2, and 3 indicated that observers have
a fairly high VSTM capacity for layout information. In
each experiment, capacity estimates, in terms of number
of stimulus elements, were significantly beyond two to
four. The effects of display size were reliable but modest
in Experiments 1 and 3, suggesting there may be a gradual
loss of fidelity with display size in some cases. However,
the stimuli in each experiment were relatively simple ele-
ments. Indeed, capacity reached its highest level with the
simplest (and identical) elements (Experiment 3).
Would layout capacity still be high if complex stimuli
are used? Our goal in Experiment 4 was to push observers
to the limit with a dense and complex naturalistic scene
containing a multitude of object identities.
The stimuli were generated from pictures of a highly
colorful dollhouse with furniture and other scene-
appropriate objects (shown in Figures 4A–4D). All stim-
uli were generated from this one dollhouse scene. Twelve
regions were defined in the scene, with each region con-
taining its own collection of identifiable, fixed-location
Table 3
Overall Percentages Correct (PC) and Estimates of Sensitivity (d′, A′), Bias (c, b), and
Capacity (k), for Each Display Size in Experiment 3
Sensitivity Bias
Capacity
Pashler
Display PC d′A′cbPashler’s k Empty Pashler Both
Size Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Irreg. Reg. Ir reg. Reg. Irreg. Reg. Irreg.
12 79.06 76.25 1.85 1.69 .86 .84 .04 .05 1.41 1.54 8.05 7.30 8.72 7.91 21.47 19.47
16 80.45 73.59 2.16 1.46 .87 .81 .07 .05 1.53 1.30 10.70 8.99 6.02 5.06 21.40 17.98
20 78.91 67.87 1.85 1.11 .86 .75 .04 .05 1.47 1.54 13.36 8.74 3.34 2.19 21.38 13.99
Note—Reg., regular; Irreg., irregular.

Sh o r t -te r M Me M o r y f o r la y o u t 1105
Pashler’s k was 5.0 stimulus regions. The SE of the
estimate was 0.28 regions. Nine of the 12 observers had
capacities greater than four regions, and the difference be-
tween 4 and the observers’ capacities was reliable [t(12) 5
3.68, p , .01]. Thus, most observers held more than four
stimulus regions in memory. If the observers represented
all of the objects in the regions (average of 2.17 per re-
gion), the total number of stimulus objects contained in
memory would be 10.8.
The location of changes varied among the three super-
regions: outer, inner-front, and inner-back (Figure 4D).
Performance varied modestly across region type, being
most accurate for changes in the outer regions (82.5%),
followed by the inner-front regions (78.0%) and the
inner- back regions (73.4%) [for the main effect of su-
perregion, F(2,22) 5 6.23, p , .001, h2
p 5 .36]. How-
ever, performance within each region type was reliably
greater than the 50% floor ( ps , .001). Thus, observers
were maintaining information about layout in each of the
four grayscale array pairs were created by converting the color pairs
to grayscale in Adobe Photoshop. Each of the resulting 108 stimulus
pairs was used twice, once for each response.
The color and grayscale stimuli were blocked in sets of 18 trials
with order counterbalanced. The type of image was illustrated before
each block. The blank interval between the first and second arrays
was 800 msec in this experiment, and the duration of the first stimu-
lus was 200 msec. Twelve students (6 female) participated.
Results
Performance decreased modestly with increasing dis-
play size, from 81.5% correct at four regions to 75.5% and
77.0% at six and eight regions, respectively (Figure 4E).
This was a total drop of 6% or 4.5% as the number of ob-
jects in the regions increased by an average of 8.4 objects.
The main effect of display size was reliable [F(2,22) 5
6.53, p , .001, h2
p 5 .37]. The d′ values were 2.14, 1.52,
and 1.78 for each respective display size (Table 4). Again,
these results are consistent with a gradual loss of fidelity
with display size.
50
60
70
80
90
100
468
AB
E
Percentage Correct
Display Size
CD
Same
Dierent
Inner Inner
back back
back
back backback
frontfront
frontfront
frontfront
OuterOuter
Figure 4. Stimuli (panels A–D) and results (panel E) from Experiment 4. Panel A
shows the dollhouse with all locations filled. Panel B shows the 12 regions. Panels C
and D show an eight- region stimulus pair in which the changing regions were the
inner-lower left-rear and the inner-lower right-front.

1106 Sa n o c k i , Se l l e r S , Mi t t e l S ta d t , a n d Su l M a n
size interaction [F(4,44) 5 7.49, p , .001] and the three-
way interaction of display size, response, and superregion
[F(4,44) 5 7.07, p , .001]. As can be seen, same re-
sponses become more likely with larger displays and less
salient critical regions (i.e., inner, but not outer, critical
regions). These changes in bias are relatively small; they
do not contravene the finding that change detection was
fairly accurate across the four factors.
Discussion
The purpose of this experiment was to measure the ca-
pacity of VSTM for layout with a very complex scene. We
found that performance was modestly affected by display
size; there was no indication of a sharp limit in short-term
memory capacity. The estimate of memory capacity was
5.0 stimulus regions, or an average of 10.8 objects (as-
suming that all objects in a region were encoded). This is
a substantial portion of a complex scene, but is still less
than the entire scene. Information was most likely to be
missing from the least salient regions (inner-back super-
region), but information was lost in all regions. This is
most consistent with the idea that there was an overall loss
in fidelity when the scene’s complexity exceeded capacity.
Complex scene details appear to have been represented at
a moderate level of resolution.
Because the scene contained some highly distinctive
objects, it is possible that the identities of objects may
have contributed to performance. For example, observ-
ers could have noticed changes in the presence of certain
objects, such as a desk or a chair. However, the high capac-
ity obtained without any object identities in Experiment 3
argues against a general explanation of VSTM capacity
that depends on object identity.
GENERAL DISCUSSION
The purpose of the present experiments was to examine
the capacity of VSTM for the layout of objects across a
sample of different stimulus situations. Consistent with
previous suggestions in the literature, we obtained esti-
mates of moderate to high capacity for layout: Capacity
estimates ranged between 5 multiobject stimulus regions
(about 11 objects), with the most complex naturalistic
scene in Experiment 4, to 13.4 stimulus elements with the
flat (2-D) grid of uniform objects in Experiment 3 (regular
condition).
VSTM capacity for layout can be illustrated in terms
of the stimulus displays and performance levels. Partici-
pants held almost all of the somewhat complex displays
three superregions. Change detection was not limited to
alterations in scene envelope, although these alterations
were most noticeable. Nor was change detection limited
to inner alterations, which would be most likely to be part
of a continuous spatial representation. This result is con-
sistent with the idea that the representation loses resolu-
tion as capacity is reached, beginning with the least salient
details (inner-back regions). However, the result can also
be explained in terms of f ixed resolution and limited spa-
tial extent if one assumes that the location of the spatial
limitation varies from trial to trial. That is, some regions
may be represented at high resolution (producing accurate
change detection), whereas other regions may not be rep-
resented. On a given trial, the least salient regions would
be the least likely to be represented.
Color images produced no advantages over grayscale
images. Accuracy was as high with grayscale images
(77.9%) as with color images (78.0%) (SED 5 1.7%).
This suggests that the potentially useful but redundant
feature of color was of no additional value in encoding
or maintaining short-term memory, at least with familiar
displays.
There were three reliable interactions in the analysis,
involving display size, superregion, and response. The
interaction of display size and response was modest in
magnitude (see Figure 4E) [F(2,22) 5 6.68, p , .001],
suggesting that there was some variation in decision bias,
with same responses becoming more likely with larger
displays. Percentages correct for all relevant conditions
are shown in Table 5. A same bias was involved in the
other two interactions as well: the superregion 3 display
Table 4
Overall Percentages Correct (PC) and Estimates of Sensitivity (d′, A′), Bias
(c, b), and Capacity (k), Averaged Over the Regularity Condition
for Each Display Size in Experiment 4
Capacity
Display Sensitivity Bias Pashler’s Pashler Pashler
Size PC d′ A′ c b k Empty Both
4 81.48 2.14 .87 2.23 1.25 2.95 5.90 8.86
6 75.52 1.52 .82 2.25 1.18 3.64 3.65 7.29
8 76.97 1.78 .84 2.37 1.48 5.00 2.51 7.50
Table 5
Percentages Correct Within Each Superregion in Experiment 4
Display Condition
Size Same Different M
Outer Regions
4 85.4 86.1 85.8
6 84.7 62.6 73.6
8 86.1 90.3 88.2
Inner-Front Regions
4 81.9 72.9 77.4
6 86.8 77.8 82.3
8 88.9 59.7 74.3
Inner-Back Regions
4 86.1 76.4 81.3
6 76.4 64.6 70.5
8 80.5 56.2 68.4

Sh o r t -te r M Me M o r y f o r la y o u t 1107
entities and their relations) within those regions. There
would be nothing special about the outer regions beyond
the ease of encoding their states and their interrelations.
Salience should also be high for the inner-front regions,
as opposed to that for the inner-back regions, because of
their increased visibility. This is reflected in the ordering
of performance in Experiment 4.
A weakness of the network model is that additional
relations can be added to account for many results, mak-
ing the model difficult to falsify. In order to deal with the
plethora of relations that may be encoded in a network,
Sanocki (1999) proposed comparing conditions in which
there should be more or fewer relations. This approach
was used in Experiment 3, in which a standard matrix with
horizontal and vertical structure was compared with an
irregular matrix. The standard matrix should support the
encoding of more relations between the elements than the
irregular matrix, and the increased memory capacity in
the standard condition is consistent with this idea.
However, relative to less systematic surfaces, texture
gradients either did not facilitate memory (Experiment 1)
or had a small positive effect (Experiment 2). Perhaps hu-
mans are fairly good at encoding interitem relations, and
the irregular gradients provided enough interrelation for
efficient encoding (cf. examples in Figures 1C and 1D and
in Figures 2C and 2D). Also, note that redundant color in-
formation may aid segmentation (e.g., Wurm et al., 1993),
but it did not aid layout memory in the present Experi-
ment 4. Additional research will be necessary to develop
the network perspective and distinguish it from other per-
spectives. Understanding interitem relations will be key to
developing this perspective.
An alternative perspective is the well-supported hy-
pothesis that VSTM capacity is defined in terms of ob-
jects and that the number of objects is limited to three
or four items (e.g., Vogel et al., 2001). In the present ex-
periments, small numbers of objects would be created
by grouping object elements into larger units. In Experi-
ments 1, 2, and 3, groups of two to four elements would
be needed to bring the total number of entities to within
four. In Experiment 4, some of the regions would have to
be grouped together to bring the number of entities down
to four. In keeping with the assumption that VSTM coding
is discrete (all or none; see, e.g., Zhang & Luck, 2008),
one could assume that each element either joins a group
or is lost and that the probability of loss increases gradu-
ally with display size. An alternative object-based model
is that attention was distributed across each stimulus array
to encode a single complex object configuration or layout
configuration.
A third perspective is that multiple encoding systems
contribute to VSTM capacity for layout. For example,
Rensink (2000a) proposed that a limited-capacity at-
tentional system for perceiving objects (i.e., VSTM) is
complemented by a larger scale system for encoding the
setting and layout of the scene. The detection of changes
in object layout could come from either the attentional
VSTM system or the layout system. The attentional sys-
tem is flexible but should excel at encoding objects and
certain interobject relations, including grouping between
of Experiment 1 in memory (Figures 1C and 1D) and held
much of the more complex displays of Experiments 2, 3,
and 4 in memory (Figures 2C, 2D, 3C, 4C, and 4D). There
was not, however, a complete representation of the object
layout in the complex displays: The intermediate levels
of performance in the later cases indicate that, although
many details were remembered, many were lost.
The numerical capacity estimates are in terms of num-
bers of stimulus elements. As noted, grouping of elements
is likely to reduce the number of representational units
below these estimates. A variety of types of grouping and
configural strategies may be employed, as discussed in
Experiment 3.
The present findings can be interpreted from several
perspectives. The experiments were motivated by the idea
that a fairly high memory capacity for layout has evolved
in humans. We propose that this ability is critical for what
can be called remote visual understanding, where the gath-
ering and integration of information from the surrounding
environment is paramount. An excellent example of re-
mote visual understanding would be the task of scouts or
watchpersons, which is to monitor surrounding areas for
activity. A high memory capacity for scenic layout would
support this duty. However, the duty is removed from im-
mediate action because the information gained need not
be used immediately. Thus, the mechanisms of remote vi-
sion and visually guided action may well be different (see
also, e.g., Goodale, Gonzalez, & Króliczak, 2008).
What can we say about the memory system that un-
derlies layout memory? We have proposed that layout
memory encodes objects and locations in a configural
manner across expanses that can be fairly broad. We be-
lieve that the memory representation is best thought of as
a hierarchical network of nodes (concepts) and the rela-
tions between them (e.g., Sanocki, 2003). At the highest
level is encoding of the scene, which would be linked to
intermediate nodes for groupings or subregions within
the scene, then the basic scene entities (e.g., objects and
surfaces). In this model, the key to memory capacity is
the encoding of relations between nodes. Relations would
include both spatial predicates, such as next to or behind,
and grouping relations, such as part of. For example, in
Experiment 4, the house scene would involve a network
that included the 12 basic scene entities (the 12 regions).
The regions would have two states: filled (e.g., with fur-
niture) or empty. Relations would encode the layout of
the regions (e.g., in front of, said of cart or dining table)
and specif ic grouping relations on a particular trial. For
example, if the left two yard regions were empty on a trial,
they could be part of the “left side (empty)” group. Filled
right-yard regions might be part of the “right side (filled)”
group. Changes of these relations, as well as of the basic
entities, could support change detection. Thus, both the
individual objects and their relations to each other are part
of the memory code (see also, e.g., Hollingworth, 2007;
Jiang et al., 2000; Olson & Marshuetz, 2005; Vidal, Gau-
chou, Tallon-Baudry, & O’Regan, 2005).
In this model, the higher accuracy of change detec-
tion for outer regions in Experiment 4 would be due to
the increased salience of the stimulus information (about

1108 Sa n o c k i , Se l l e r S , Mi t t e l S ta d t , a n d Su l M a n
of Experimental Psychology: Human Perception & Performance, 30,
519-537.
Hollingworth, A. (2005). The relationship between online visual rep-
resentation of a scene and long-term scene memory. Journal of Exper-
imental Psychology: Learning, Memory, & Cognition, 31, 396-411.
Hollingworth, A. (2006). Scene and position specificity in visual
memory for objects. Journal of Experimental Psychology: Learning,
Memory, & Cognition, 32, 58-69.
Hollingworth, A. (2007). Object-position binding in visual memory
for natural scenes and object arrays. Journal of Experimental Psychol-
ogy: Human Perception & Performance, 33, 31-47.
Hollingworth, A., Hyun, J.-S., & Zhang, W. (2005). The role of
visual short-term memory in empty cell localization. Perception &
Psychophysics, 67, 1332-1343.
Intraub, H. (1997). The representation of visual scenes. Trends in Cog-
nitive Sciences, 1, 217-222.
Irwin, D. E. (1991). Information integration across saccadic eye move-
ments. Cognitive Psychology, 23, 420-456.
Irwin, D. E., & Zelinsky, G. J. (2002). Eye movements and scene per-
ception: Memory for things observed. Perception & Psychophysics,
64, 882-895.
Jiang, Y., Olson, I. R., & Chun, M. M. (2000). Organization of visual
short-term memory. Journal of Experimental Psychology: Learning,
Memory, & Cognition, 26, 683-702.
Lasaga, M. I., & Garner, W. R. (1983). Effect of line orientation on
various information-processing tasks. Journal of Experimental Psy-
chology: Human Perception & Performance, 9, 215-225.
Liu, K., & Jiang, Y. (2005). Visual working memory for briefly pre-
sented scenes. Journal of Vision, 5, 650-658.
Logie, R. H. (1995). Visuo–spatial working memory. Hillsdale, NJ:
Erlbaum.
Luck, S. J. (2008). Visual short-term memory. In S. J. Luck & A. Hol-
lingworth (Eds.), Visual memory (pp. 43-85). New York: Oxford Uni-
versity Press.
Melcher, D. (2001). Persistence of visual memory for scenes. Nature,
412, 401.
Melcher, D. (2006). Accumulation and persistence of memory for
natural scenes. Journal of Vision, 6, 8-17.
Miller, G. A. (1956). The magical number seven, plus or minus two:
Some limits on our capacity for processing information. Psychologi-
cal Review, 63, 81-97.
Ni, R., Braunstein, M. L., & Andersen, G. J. (2005). Distance percep-
tion from motion parallax and ground contact. Visual Cognition, 12,
1235-1254.
Olson, I. R., & Marshuetz, C. (2005). Remembering “what” brings
along “where” in visual working memory. Perception & Psychophys-
ics, 67, 185-194.
O’Regan, J. K. (1992). Solving the “real” mysteries of visual percep-
tion: The world as an outside memory. Canadian Journal of Psychol-
ogy, 46, 461-488.
Palmer, S. E. (1975). Visual perception and world knowledge: Notes
on a model of sensory–cognitive interaction. In D. A. Norman, D. E.
Rumelhart, & LNR Research Group (Eds.), Explorations in cognition
(pp. 279-307). San Francisco: Freeman.
Pashler, H. (1988). Familiarity and visual change detection. Perception
& Psychophysics, 44, 369-378.
Phillips, W. A. (1974). On the distinction between sensory storage and
short-term visual memory. Perception & Psychophysics, 16, 283-290.
Rensink, R. A. (2000a). The dynamic representation of scenes. Visual
Cognition, 7, 17-42.
Rensink, R. A. (2000b). Visual search for change: A probe into the na-
ture of attentional processing. Visual Cognition, 7, 345-376.
Sanocki, T. (1999). Constructing structural descriptions. Visual Cogni-
tion, 6, 299-318.
Sanocki, T. (2003). Representation and perception of scenic layout.
Cognitive Psychology, 47, 43-86.
Sanocki, T., Michelet, K., Sellers, E., & Reynolds, J. (2006). Rep-
resentations of scene layout can consist of independent, functional
pieces. Perception & Psychophysics, 68, 415-427.
Sanocki, T., & Sulman, N. (2009). Priming of simple and complex
scene layout: Rapid function from the intermediate level. Journal of
nearby objects. In contrast, the layout system could encode
the arrangement of major entities in the scene, as well as
scenewide properties, such as the scene envelope. Thus,
in this model, the increased accuracy for outer changes
in Experiment 4 could be attributed to information about
the scene envelope from the layout system. Inner-front
regions are probably more likely to be attended than inner-
back regions, and this would explain the increased accu-
racy for inner-front regions.
One other feature of the multiple-system model is
that there is not a single short-term memory system for
scene layout information. Whereas the attentional system
may be short term in nature, the layout system could be a
working memory that serves as an entrance to long-term
memory (see Melcher, 2001, 2006).
AUTHOR NOTE
Portions of the present results were presented at the First Annual
Meeting of the Vision Sciences Society, May 2001, and the 42nd Annual
Meeting of the Psychonomic Society, November 2001. We thank mem-
bers of the Vision–Cognition–Design Lab at the University of South
Florida for their help in conducting the experiments. Correspondence
concerning this article should be addressed to T. Sanocki, Department of
Psychology, PCD 4118, University of South Florida, Tampa, FL 33620-
8200 (e-mail: sanocki@usf.edu).
REFERENCES
Aginsky, V., & Tarr, M. J. (2000). How are different properties of a
scene encoded in visual memory? Visual Cognition, 7, 147-162.
Alvarez, G. A., & Cavanagh, P. (2004). The capacity of visual short-
term memory is set both by visual information load and by number of
objects. Psychological Science, 15, 106-111.
Alvarez, G. [A.], & Oliva, A. (2007). The role of global layout in vi-
sual short-term memory. Visual Cognition, 15, 70-73.
Awh, E., Barton, B., & Vogel, E. K. (2007). Visual working memory
represents a fixed number of items regardless of complexity. Psycho-
logical Science, 18, 622-628.
Baddeley, A. D., & Hitch, G. J. (1974). Working memory. In G. H.
Bower (Ed.), The psychology of learning and motivation: Advances in
research and theory (Vol. 8, pp. 47-89). New York: Academic Press.
Brockmole, J. R., Wang, R. F., & Irwin, D. E. (2002). Temporal
integration between visual images and visual percepts. Journal of
Experimental Psychology: Human Perception & Performance, 28,
315-334.
Eng, H. Y., Chen, D., & Jiang, Y. (2005). Visual working memory for
simple and complex visual stimuli. Psychonomic Bulletin & Review,
12, 1127-1133.
Enns, J. T. (1987). A developmental look at pattern symmetry in percep-
tion and memory. Developmental Psychology, 23, 839-850.
Franconeri, S. L., Alvarez, G. A., & Enns, J. T. (2007). How many
locations can be selected at once? Journal of Experimental Psychol-
ogy: Human Perception & Performance, 33, 1003-1012.
Gibson, J. J. (1979). The ecological approach to visual perception. Bos-
ton: Houghton Mifflin Harcourt.
Goodale, M. A., Gonzalez, C. L. R., & Króliczak, G. (2008). Action
rules: Why the visual control of reaching and grasping is not always
influenced by perceptual illusions. Perception, 37, 355-366.
Grimes, J. (1996). On the failure to detect changes in scenes across
saccades. In K. Akins (Ed.), Perception (Vancouver Studies in Cog-
nitive Science, Vol. 5, pp. 89-110). New York: Oxford University
Press.
He, Z. J., Wu, B., Ooi, T. L., Yarbrough, G., & Wu, J. (2004). Judging
egocentric distance on the ground: Occlusion and surface integration.
Perception, 33, 789-806.
Hollingworth, A. (2004). Constructing visual representations of natu-
ral scenes: The roles of short- and long-term visual memory. Journal

Sh o r t -te r M Me M o r y f o r la y o u t 1109
tures, conjunctions, and objects in visual working memory. Journal
of Experimental Psychology: Human Perception & Performance, 27,
92-114.
White, L. E., Coppola, D. M., & Fitzpatrick, D. (2001). The contri-
bution of sensory experience to the maturation of orientation selectiv-
ity in ferret visual cortex. Nature, 411, 1049-1052.
Wolfe, J. M., Klempen, N., & Dahlen, K. (2000). Postattentive vision.
Journal of Experimental Psychology: Human Perception & Perfor-
mance, 26, 693-705.
Wurm, L. H., Legge, G. E., Isenberg, L. M., & Luebker, A. (1993).
Color improves object recognition in normal and low vision. Journal
of Experimental Psychology: Human Perception & Performance, 19,
899-911.
Zhang, W., & Luck, S. J. (2008). Discrete fixed-resolution representa-
tions in visual working memory. Nature, 453, 233-235.
(Manuscript received June 4, 2008;
revision accepted for publication January 18, 2010.)
Experimental Psychology: Human Perception & Performance, 35,
735-749.
Sebrechts, M. M., & Garner, W. R. (1981). Stimulus-specific pro-
cessing consequences of pattern goodness. Memory & Cognition, 9,
41-49.
Simons, D. J. (1996). In sight, out of mind: When object representations
fail. Psychological Science, 7, 301-305.
Torralba, A. (2003). Contextual priming for object detection. Interna-
tional Journal of Computer Vision, 53, 169-191.
Tsal, Y., Meiran, N., & Lamy, D. (1995). Towards a resolution theory
of visual attention. Visual Cognition, 2, 313-330.
Vecera, S. P., & Palmer, S. E. (2006). Grounding the figure: Surface
attachment influences f igure–ground organization. Psychonomic Bul-
letin & Review, 13, 563-569.
Vidal, J. R., Gauchou, H. L., Tallon-Baudry, C., & O’Regan, J. K.
(2005). Relational information in visual short-term memory: The
structural gist. Journal of Vision, 5, 244-256.
Vogel, E. K., Woodman, G. F., & Luck, S. J. (2001). Storage of fea-