Cognitive maps are mental models of the relative
locations and attributes of phenomena in spatial environments.
Understanding how people form cognitive maps of virtual environments
is vital to effective virtual world design. Unfortunately, such
an understanding is hampered by the difficulty of cognitive map
measurement. The present study tests the validity of using sketch
maps to examine aspects of virtual world cognitive maps. We predict
that subjects who report feeling oriented within the virtual world
will produce better sketch maps and so sketch map accuracy can
be used as an external measure of subject orientation and world
knowledge. Results show a high positive correlation between subjective
ratings of orientation, world knowledge and sketch map accuracy,
supporting our hypothesis that sketch maps provide a valid measure
of internal cognitive maps of virtual environments. Results across
different worlds also suggest that sketch maps can be used to
find an absolute measure for goodness of world design.
KEYWORDS Cognitive Mapping, Virtual Environments,
Sketch Maps, Mental Models.
Whether in real or virtual space we form cognitive
maps to deal with and process the information contained in the
surrounding environment. Cognitive mapping is formally defined
by Downs and Stea [6] as:
"..a process composed of a series of psychological
transformations by which an individual acquires, codes, stores,
recalls, and decodes information about the relative locations
and attributes of phenomena in their everyday spatial environment."
An individual's cognitive map is an active
information seeking structure of which spatial imagery is but
one aspect [14]. Cognitive maps are also made up of memories of
objects and kinesthetic, visual and auditory cues [8].
The fundamental importance of an effective
cognitive map is that it allows two questions to be answered quickly
and efficiently: Where is that? How do I get to there from here?
Thus human spatial behavior relies upon and is determined by the
individual's cognitive map of the surrounding environment. In
addition, the perception of the environment itself is always guided by some
sort of cognitive map, so an inaccurate or incomplete cognitive
map leads to disorientation and confusion[l4].
Designing virtual worlds through which subjects
can navigate and orientate themselves successfully requires an
understanding of cognitive map formation in virtual environments.
Considerable research which might be brought to bear on this topic
has been conducted on the development of cognitive maps and how
they affect real world behavior.
In exploring how people formed mental images
of a city Briggs[4] has identified three complementary ways in
which cognitive maps are created:
Of these, an individual's sensory modalities
provide direct sources of information and are more effective in
cognitive map formation than indirect sources[6].
Cognitive maps are created as the result of
active and passive modes of information processing [14]. Generally,
active information processing gives the greatest meaning to the
information processed and produces more information for the moving
perceiver. Thus the information produced by locomotion is fundamental
to an individual's spatial orientation.
An individual's cognition of the environment
is not only a function of the behavior by which information is
obtained but also depends on the characteristics of the environment
[4]. The amount of information gained by each sensory modality
is also environmentally dependent [16].
Aside from the way cognitive maps are formed, the types of information stored in a cognitive map are also of interest. Kuipers[10] suggests that a cognitive map consists of five different types of information, each with it's own representation: Topological, Metric, Route Descriptions, Fixed Features and Sensory Images. Different techniques are needed to measure each different information type.
Finally, Lynch[12] notes the uniquely personal
nature of cognitive maps. Across different cultures he found that
different groups may have widely different images of the same
outer reality. Also, on an individual level, what an observer
sees is based on a common exterior form, but how the observer
interprets and organizes this form is unique. This interpretation
governs how the observer directs his attention and this in turn
affects what is seen. So at both a societal level and a cultural
level cognitive maps are highly individualistic.
COGNITIVE MAPPING - THE VIRTUAL EXPERIENCE
As suggested above, cognitive maps are most
effectively formed by active interaction with the environment
using many different sensory modalities. However, in a virtual
world there is typically sensory degradation and a lack of many
of the perceptual cues used in the real world. Downs and Stea
[6] point out that any filtering of information before it reaches
the sensory modalities affects the cognitive map. This is the
case for virtual environments. For example, the visual modality
may suffer from low image resolution, poor image quality or a
reduction of the peripheral field. In real environments, Alfano
and Michel[1] have shown the reduction of peripheral vision impairs
perception and visuomotor performance, both of which are essential
for cognitive mapping ability. In addition there are rarely any
tactile or olfactory cues and often only limited auditory feedback.
The study presented here examines some of the factors influencing
cognitive map construction given current immersive technology.
METHODS FOR ASSESSING COGNITIVE MAPS
One of the difficulties in studying cognitive
mapping is the problem of extracting an external representation
of an individual's internal map. By definition a cognitive map
is highly subject-specific and, although individuals often record
the same things in their cognitive maps, there is no evidence
that they record them in the same way. Golledge[7] identifies
four distinct methods for extracting environmental cognition information:
We are particularly interested in the subject's
topological understanding of the virtual environment, i.e. knowing
where they are and where everything else is, as compared with
metric knowledge - knowing precise object location and distance
between objects. Topological knowledge is generally more important
than metric knowledge for effective navigation.
A common approach for measuring topological
knowledge was suggested by Lynch[12], who had subjects sketch
maps to represent the mental models of their local
cities. Lynch finds that sketch maps are more accurate when used
for topological rather than metric analysis.
Golledge[7] points out, however, that caution
must be taken that sketch maps are not over analyzed. The disadvantages
of sketch maps include trying to represent a three-dimensional
cognitive map in two dimensions and the difficulties of quantitative
analysis. They may also measure more than just spatial understanding
of an environment, e.g. drawing or memory ability. Conversely,
Blades[3] finds them reliable over time and Newcombe[15] comments
that they are no less accurate than other cognitive techniques.
Other common techniques used for cognitive
map analysis include distance and angle estimation. However, Henry[9]
found that distances were consistently underestimated in virtual
environments and that angle estimation produced wildly varying
results. Moreover, his subjects' sketch maps are topologically
accurate even when the sketched distances are not. In a prior
work we used a different technique for distance estimation and
found similar results[l8].
The present study is designed to asses the
validity of sketch maps as a tool for measuring cognitive maps
of virtual environments, particularly the topological knowledge
of the cognitive maps. We predict that subjects who report feeling
oriented and unconfused in the virtual world will later produce
relatively accurate sketch maps, whereas subjects who report feeling
disoriented and confused in the virtual world will produce less
accurate sketch maps. In other words, if sketch maps are an accurate
external representation of the subject's cognitive map then we
would expect a correlation between the sketch map scores and subjective
ratings of how oriented subjects felt within the virtual world.
Eighty four subjects experienced a number of
simple virtual worlds and then produced maps. The worlds were
constructed using Swivel and Body Electric software, and rendered
on an SGI VGX. Participants wore VPL Eyephones and interacted
with the virtual environment using a VPL Dataglove. Movement through
the virtual environment was achieved by the users pointing in
the desired direction they and making a "fly" gesture
with the Dataglove. This movement was completely unconstrained
so participants could be as close or far away from the world as
they wanted. Collision detection was not used so participants
could travel through objects.
Each subject was initially trained on the same immersive virtual environment until they felt comfortable with moving and interacting within a virtual environment. Following this training, they were given a 24-question survey which asked for responses on a range of navigation, orientation, interaction, presence and interface questions. Survey responses were indicated on a 10-point anchored scale. These survey questions are reproduced in the appendix. Participants were also invited to comment about the experience in general.
Table 1.0: The Different Characteristics of
the Three Test Worlds.
After the training session. subjects experienced
one of three different virtual worlds for 10 minutes and were
told to explore it as fully as they could. They were then asked
to produce a map of the world that someone unfamiliar with the
world could use to navigate around the world. The subjects also
completed the same survey that was administered after the training
world and were video taped for later observation of behavior patterns.
If the sketch maps are an accurate external
representation of the subjects cognitive map then we would expect
a correlation between the sketch map scores and subject survey
scores for orientation within the virtual world.
World Differences
Three different worlds were used to explore
how differences in world design might affect the cognitive map
formed and the resultant sketch maps. According to Darken and
Silben's[5] world classification, each of them are "small",
in that all of the world can be seen from a single viewpoint.
They are also static, all their objects having positions and values
which don't change over time. However, the density of each of
the worlds varied considerably as detailed below. Each subject
experienced only one of the test worlds.
Virtual Valley
Under Darken and Silben's scheme this is a
dense world: it has a large number of objects and spatial cues;
however, they are all placed in a logical manner. The world
is bound on either side by tall mountain ranges that direct attention
to the objects contained in the valley below. Objects within the
world are all representative of what would be expected in a real
valley and there are no hidden objects. Objects are clearly distinguishable
by color and size, and there are a number of distinctive objects
that could serve as landmarks. This world design would make it
difficult for subjects to become disoriented.
Cloudlands
Cloudlands is a sparse world containing few
objects. It contains a dominant ground plane with clusters of
objects floating above it in cloud groups. One of these clouds
contains a fish and star while the others are empty. The objects
are incongruous and surprising - there is a floating cactus, stacks
of multicolored planes, cones and small gray rocks. The are no
environmental cues to direct attention other than the object clusters
themselves. However, the sparcity of the world would also make
it difficult for subjects to become disoriented.
Neighborhood
Neighborhood is a cluttered world containing
clusters of buildings all closely grouped and each containing
other objects. The buildings are largely the same
size and color making it hard to distinguish between them, and
the objects within them are almost all the same color as the buildings.
The objects are all those that would be logically found in a neighborhood,
such as trees, tables, glass and a piano but the similarity of
the buildings makes it hard to precisely locate them. This world
is generally confusing and disorientating.

Table 1.0 summarizes the characteristics of
the three test virtual worlds.
As mentioned before, one of the challenges
of using sketch maps is analyzing the results. The maps produced
are as individualistic as each of the cognitive maps of the subjects.
Although sketch maps are commonly used in real world cognitive
mapping there is no generally accepted method for their analysis.
Useful approaches have been reported in Appleyard[2] Ladd[11],
Moore[13], and Walsh et. al.[l7], among others; however
these are used to analyze maps of large scale urban environments.
Adapting these methods, we use a simple, purely topological technique.
Each sketch map was given a set of goodness, object class and
object positioning scores as detailed below:
Map Goodness
Maps were ranked for goodness on a scale of
1-3 by two researchers who were experienced in virtual environments
but blind to subject identity and other correlated measures. The
researchers were told to rank the maps on how useful they would
be as a navigational tool if they were taken with them into the
virtual environment. They were told to ignore the participants
drawing ability and focus on how well the map represented the
virtual world and the locations of the objects within it.
Object Classes
Each map was given a score according to the
number of object classes present - for example, trees, rocks and
mountains are each counted as separate classes. Using object classes
is a way to assess completeness of a sketch map for a given virtual
world.
Relative Object Positioning
To provide a measure of differences in cognitive maps for different worlds we scored maps according to relative object positioning. We used topological positioning and so scored objects if they were correctly positioned to the right or left, above or below, or clockwise or counterclockwise, depending on the specific world being represented. The specific object position was not important, only its position relative to other objects in the sketch map.
Maps were given two positioning scores: a total
object position score in which all the objects were scored, and
a significant object position score where the five most commonly
drawn objects for each world are scored. Relative object positioning
is a way to assess the accuracy of sketch maps.
Although subjects were given no instructions
on how to produce their maps, almost all of them drew three dimensional
representations of the virtual world. This may be due to the small
size of the worlds - sketch maps produced of large scale real
world environments are usually two dimensional. Figure 1.0 shows
typical sketch maps produced for the Cloudlands world.

Table 2.0: Goodness and Class Number
correlations with virtual world orientation and knowledge
across the test worlds.
Within World Correlation
If sketch maps can be used as an external measure
of subjects' cognitive maps then there should be a strong correlation
between map goodness scores and subject scores for the survey
questions "Knowing where everything is in the Virtual World"
and "Orientation in the Virtual World". To investigate
this we correlated the object class and map goodness scores with
the survey responses. Table 2.0 shows the correlation values of
the map scores and survey scores. Although the map goodness rankings
are highly subjective, the correlation between the scores given
by the two researchers was very high; ( r = 0.86, 0.71, 0.70,
for three worlds respectively, significant at p < 0.01).
In the Virtual Valley and Neighborhood worlds
object class and map goodness were both significantly correlated
with the subjects' reported sense of orientation in the virtual
world. For these two worlds, the map goodness score is also significantly
correlated with subjects' knowledge of where everything is. However,
this isn't the case with the Cloudlands world. The sparse nature
of Cloudlands may make it difficult to produce an accurate sketch
map. Cloudlands was also more three-dimensional that the other
worlds with most objects placed high above the dominant ground
plane, adding to the difficulty of producing a two-dimensional
representation.
Since cognitive maps are most effectively formed
by active interaction with the environment, there should also
be a relationship between map scores and the survey
questions relating to interaction. This is indeed the case with
Virtual Valley, where the map goodness rankings correlate significantly
with the subjects survey score for ease of interaction (r = 0.882),
ease of navigation (r = 0.865), ease of movement within the virtual
world (r = 0.814) and ease of use of the Data Glove (r = 0.645).
However, in the other two worlds the correlation between these
survey questions and the map rankings were not significant.
Between World Differences
A two factor ANOVA was done on the survey results
to identify world differences and possible gender-linked factors.
There was a significant difference between worlds in subject's
understanding of where everything was (F[2,22]=4.49, p < 0.025),
and how oriented the subjects felt within each of the worlds (F[2,22]=3.314,
p < 0.05). For both of these questions subjects rated Neighborhood
world significantly lower than the two other worlds, as shown
in figure 2.0. There was also a significant difference between
the sense of dizziness reported by subjects, with those in Neighborhood
registering the most dizziness, (F[2,22] = 3.95, p < 0.025).
These results reflect the particularly disorienting nature of
Neighborhood world.
If sketch maps are representative of subjects
virtual world cognitive maps, they should also reflect these world
differences. The relative object position scores can be used to
compare across worlds. For each world we defined the five most
commonly drawn objects as "significant objects" and
a relative positioning ratio was then calculated for each map:
Ratio = Correctly placed significant objects divided by Total number of significant objects in map.
An ANOVA revealed a statistically significant
world difference for the significant object relative positioning
ratios, (F[2,22] = 4.004, p < 0.025). A similar ratio was calculated
for the relative positioning for all objects drawn in the sketch
maps. In this case an ANOVA showed no significant world difference,
(F[2,22] < 1.0 NS). figure 3.0 shows the relative positioning
ratios for both sets of objects.
In Virtual Valley over 90% of the significant
objects that are placed are placed correctly, reflecting the well
designed nature of the world. The difference in ratios from "significant" object placement to "all" object placement in Virtual
Valley is largely due to a number of landmark objects which almost
all of the subjects positioned correctly. The similarity of the
"significantquot; and "all" object placement ratios
in the other worlds may mean that there are fewer, if any, landmark
objects.
The difference in Virtual Valley and Neighborhood
map scores correspond to the difference in subjects' orientation
scores shown in figure 2.0. This suggests that "significant
object- positioning scores may be used as a simple absolute measure
of map accuracy and goodness of world design. It also implies
that the sketch maps for these worlds accurately represents the
topological knowledge stored in the subjects' cognitive maps.
Figure 3.0: Average subject relative object
positioning ratios across the three test worlds.
In this study we have investigated the applicability
of sketch maps as an external representation of an individual's
cognitive map of a virtual environment. We have found that sketch
maps reflect differences both between worlds and within worlds.
We used three methods to score the sketch maps,
chosen for their simplicity and general applicability: map goodness
and object class number for comparing maps from a given world,
and the relative object positioning ratio for comparing maps across
a range of worlds.
In two of our test worlds, Virtual Valley and
Neighborhood, map goodness and object class number scores correlated
significantly with the subjects' self-reported sense of orientation
within the virtual world. The relative object positioning ratio
also matched the difference in reported orientation between Virtual
Valley and Neighborhood worlds. These two results suggest that
sketch maps do indeed accurately represent the topological aspects
of subjects cognitive maps.
The "significant object" ratio appears
useful for comparing across worlds, while the map goodness and
object class scores are useful for comparing subjects within worlds.
The difference between the "significant" and "all"
object placement ratios may also be used to identify worlds that
have well defined landmarks.
However, the low correlation with the Cloudlands
results may indicate that sketch mapping is more useful for relatively
dense worlds, or that more complicated forms of sketch map analysis
is needed for sparse worlds.
ACKNOWLEDGMENTS
The authors would like to thanK the reviewers
for their useful comments, and Hunter Hoffman and William Winn
for help with analysis of the experimental results.
REFERENCES
1.Alfano, P.L. and Michel, G.F. Restricting the Field of View: Perceptual
and Performance Effects. Perceptual and Motor Skills, 1990,
70,pp3545.
2. Appleyard, D. Styles and Methods of Structuring
a City. Environment and Behavior, 1970, vol 2, pp 100-118.
3. Blades. M. The Reliability of Data Collected
From Sketch Maps. Journal of Environmental Psychology, 1990,
vol 10, pp 327-339.
4. Briggs, R. Urban Cognitive Distance. In
Image and Environment, (Downs, R.M. and Stea, D. Eds.),
Aldine Publishing Co., Chicago, 1973,pp361-388.
5. Darken, R.P. and Silbert, J.L. A Toolset
for Navigation in Virtual Environments. In Proceedings of the
ACM Symposium on User Interface Software and
Technology, UIST'93, 1993, pp l57-166.
6. Downs, R.M. and Stea, D. Cognitive Maps
and Spatial Behavior. Process and Products. In Image and Environment,
(Downs, R.M. and Stea, D. Eds.), Aldine Publishing Co., Chicago,
1973, pp 8-26.
7. Golledge, R.G. Methods and Methodological
Issues in Environmental Cognition Research. In Environmental
Knowing, (Golledge, R.G. and Moore, G.T. Eds.), Dowden, Hutchinson
and Ross, Inc., Pennsylvania, 1976, pp 300-313.
8. Griffin, D.R. Topographical Orientation.
In Image and Environment, (Downs, R.M. and Stea, D. Eds.),
Aldine Publishing Co., Chicago, 1973, pp 296-299.
9. Henry, D. Spatial Perception in Virtual
Environments: Evaluating an Architectural Application. M.S.E.
Thesis, College of Engineering, University of Washington, 1992.
10. Kuipers, B. The Cognitive Map: Could it
have been any other way. In Spatial Orientation: Theory, Research
and Application, (Pick, H.L. and Acredolo, L.P. Eds.), Plenium
Press, New York, 1983, pp 345-360.
11. Ladd, F.C. Black Youths View their Environment:
Neighborhood Maps. Environmental Behavior, 1970, vol 2,
pp 64-79.
12. Lynch, K. The Image of the City, MIT
Press, Cambridge, Massachusetts, 1960.
13. Moore, G.T. Theory of Research and the
Development of Environmental Knowing. In Environmental Knowing,
(Golledge, R.G. and Moore, G.T. Eds.), Dowden, Hutchinson
and Ross, Inc., Pennsylvania, 1976, pp l38-164.
14. Neisser, U. Cognition and Reality, WH
Freeman, San Francisco, 1976.
15. Newcombe, N. Methods for the Study of Spatial
Cognition. In The Development of Spatial Cognition, (Cohen,
R. Ed.), Hillsdale, New Jersey, Lawrence Erlbaum Associates, 1985,
pp 277-300.
16. Stea, D. and Blaut, J.M. Some Preliminary
Observations on Spatial Learning in School Children. In Image
and Environment, (Downs, R.M. and Stea, D. Eds.), Aldine Publishing
Co., Chicago, 1973, pp 226-234.
17. Walsh, D.A., Krauss, IK., and Reginer,V.A.
Spatial Ability, Environmental Knowledge, and Environmental Use:
The Elderly. In Spatial Representation and Behavior Across
the Life Span: Theory and Application, (Liben, L.S., Patterson,
A.H. and Newcombe, N. Eds.), Academic Press, New York, 1981, pp
321-357.
18. Weghorst, S. and Billinghurst, M. Spatial
Perception of Immersive Virtual Environments, HIT Lab Technical
Report, University of Washington, 1993.
The 24 survey questions given to subjects are
listed below. For each of the questions subjects were asked to
rank their responses on a scale from one to ten. The anchors for
these scales are shown under the each of the questions. Responses
were collected automatically using a Hypercard stack on a Macintosh
computer and participants were also given the opportunity to add
their own comments at the end of the survey.
Questions
1. Sense of being there:
None - Total
2. Ease of interaction:
Impossible - Effortless
3. Comfort of the display hardware:
Unbearable - Comfortable
4. Enjoyment:
Boring - Very enjoyable
5. How easy was it to navigate?
Very difficult - Very easy
6. Sense of orientation relative to the laboratory:
No sense of direction - Completely orientated
7. Sense of orientation in the virtual world:
No sense of direction - totally orientated
8. Feeling of being lost:
All the time - Never
9. Sense of dizziness:
Never - All the time
10. Image brightness:
Way too dim - Way too bright
11. Color quality:
Very poor - Very good
12. Ease of use of the glove:
Very difficult - Very easy
13. Feeling of inclusion in the world:
Totally removed - Actually there
14. Overall physical comfort:
Very uncomfortable - Very comfortable
15. Understanding of where everything was in
the world
Total confusion - Total understanding
16. Invites exploration:
Not at all - Very much so
17. Invites introspection:
Not at all - Very much so
18. Ease of movement around the world:
Very difficult - Very easy
19. Ease of getting where you wanted to go:
Very easy - Impossible
20. How engaging was it?
Not at all - Totally
21. Image clarity:
Extremely fuzzy - Extremely sharp
22. How comfortable are you with using computers?
Totally uncomfortable - Totally comfortable
23. Your experience in Virtual Reality:
First time - Very Experienced
24. Sense of presence within the Virtual World:
Very low - Very High
Figure 1.0: Cloudlands world (upper left) and
three typical sketch maps.
Figure 2.0 Average subject orientation and world knowledge survey scores across the three test worlds
