Goldberg Visual Scanpath Representation

Visual Scanpath Representation
Joseph H. Goldberg and Jonathan I. Helfman
Applications User Experience
Oracle USA
joe.goldberg@oracle.com; jon.helfman@oracle.com

Abstract attention at particular positions, and saccades, rapid eye move-
ments with suppressed visual perception. Fixation positions are
Eye tracking scanpaths contain information about how people specified with respect to the background stimulus image or dis-
see, but traditional tangled, overlapping scanpath representations play that was visible when the samples were collected. A se-
provide little insight about scanning strategies. The present work quence of consecutive fixations and saccades results in a scan-
describes and extends several compact visual scanpath represen- path – a trace through time and space that may overlap itself.
tations that can provide additional insight about individual and Scanpaths, along with other ocular-based metrics, may indicate
aggregate/multiple scanning strategies. Three categories of rep- higher cognitive strategies and states [e.g., Marshall 2007; Hor-
resentations are introduced: (1) Scaled traces are small images of nof and Halverson 2003]. A challenge for the researcher is to
scanpaths as connected saccades, allowing the comparison of compare multiple scanpaths and quickly comprehend each par-
relative fixation densities and distributions of saccades. (2) Time ticipant’s scanning strategy in each tested condition.
expansions, substituting ordinal position for either the scan-
path’s x or y-coordinates, can uncover otherwise subtle horizon- Typical visual representations of scanpaths use circles to repre-
tal or vertical reversals in visual scanning. (3) Radial plots represent fixations and lines to represent saccades. Figure 1 displays a
sent scanpaths as a set of radial arms about an origin, with each hypothetical scanpath with 5 fixations and 4 saccades. A scan-
arm representing saccade counts or lengths within a binned set path may be replayed using animation, but to show a scanpath as
of absolute or relative angles. Radial plots can convey useful a static image, fixation duration is typically coded by circle ra-
shape characteristics of scanpaths, and can provide a basis for dius.
new metrics. Nine different prototype scanning strategies were
represented by these plots, then heuristics were developed to
classify the major strategies. The heuristics were subsequently
applied to real scanpath data, to identify strategy trends. Future
work will further automate the identification of scanning strate-
gies to provide researchers with a tool to uncover and diagnose
scanning-related challenges.

CR Categories: H.1.2. User/Machine Systems: Software Psy-
chology, H.5.2. User Interfaces: Graphical User Interfaces. Figure 1. A hypothetical scanpath with 5 fixations and 4 sac-
cades, showing fixation order and varying saccade lengths.
Keywords: Eye Tracking, Usability Evaluation, Scanpath,
Scanning Strategy, Visualization Actual scanpath records are usually quite complex, and can be
difficult to interpret and compare. Figure 2 shows a scanpath
from a single observer trying to locate content within the lower
1 Introduction right hand area of a web page. The scanpath visited multiple
content columns as well as the search area before locating the
It has been said that proper representation of a problem is the content. Frequent crossing of saccades and revisiting of loca-
single most important step towards its eventual solution. The tions contribute to scanpath complexity.
problem we are investigating is the comparison of visual scan-
ning strategies between individuals, groups, and conditions, in 1.2 Scanpath Comparison
order to make inferences for improving the visual design of
software and digital media. This paper introduces and extends Numerical techniques have been presented for summarizing and
several representations for scanpaths, then demonstrates how comparing scanpaths.. The length of a scanpath can be a meas-
these representations can be used to develop heuristics and met- ure of productivity, and average fixation duration may measure
rics for detecting and labeling scanning strategies. of cognitive complexity [Goldberg and Kotval 1999]. The sum
and distribution of the inter-fixation angles within a scanpath
1.1 Scanpaths have been proposed as a measure of efficiency on a task, with
simple, direct scanpaths signaling higher efficiency [Goldberg
Eye tracking methods typically sample gaze locations at 50-120 and Kotval 1998]. String editing distance computes the mini-
Hz. Samples are then reduced to fixations, periods of visual mum number of editing steps required to transform one scanpath
Copyright © 2010 by the Association for Computing Machinery, Inc.
into another [West, et al. 2006]. Sequence alignment techniques
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed first attempt to align sequences as closely as possible, before
for commercial advantage and that copies bear this notice and the full citation on the computing their disimilarity [Josephson and Holmes 2002].
first page. Copyrights for components of this work owned by others than ACM must be Matching alignments from multiple sequences may also repre-
honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on
sent the ‘averaged’ scanpath from a set of users [Hembrooke, et
servers, or to redistribute to lists, requires prior specific permission and/or a fee.
Request permissions from Permissions Dept, ACM Inc., fax +1 (212) 869-0481 or e-mail
permissions@acm.org.
ETRA 2010, Austin, TX, March 22 – 24, 2010.
© 2010 ACM 978-1-60558-994-7/10/0003 $10.00
203

2.1 Scaled Traces

A scaled trace is a spatial representation of a scanpath sequence
that shows saccadic lengths and inter-fixation angles, but not
fixation durations. In classic studies of the influence of intent on
eye movements, Yarbus [1967] demonstrated that eye move-
ment scanpaths gathered over 3 minute periods differed dramati-
cally depending on the demands of specific tasks. Our represen-
tation of scaled traces, derived from this work, is shown in Fig-
ure 3.

Figure 2. A scanpath from a single observer trying to locate
content within the lower right hand area of a web page.
Figure 3. Representation of scanpath (left) as a trace image
al., 2006]. Statistical tests are also available for testing signifi- (right) by excluding fixations and minimizing scale.
cant differences between sets of scanpaths. In one comparison
approach, Heminghous and Duchowski [2006] provide a method In addition to showing a standard trace line (Figure 4A), sequen-
to automate the process of assigning areas of interest. They first tial fixation order can also be represented by changing graphics
cluster scanpaths using a mean shift algorithm [Santella and properties of the line as it is drawn from beginning to end, from
DeCarlo, 2004], then tabulate similarity coefficients within and light to dark (Figure 4B), or from thin to thick (Figure 4D).
between participants and images. In another approach, Feusner Multicolored segments can also be used (Figure 4C), but it may
and Lukoff [2008] compute the distance between all pairwise be more difficult to convey fixation order with color than with
combinations of scanpath sequences using editing distance, then brightness or thickness.
determine the difference between the average distances within
and between groups to be compared. Using Monte Carlo simula-
tion methods, they generate significance values for observed
differences.

Scanpaths can also be summarized and compared visually. This
paper provides a set of compact visual representations of scan-
paths that can help usability professionals identify and compare
scanning strategies in eye-tracking data. Visual representations
can provide rich summaries of scanpath metrics, such as the
distribution of inter-fixation angles and saccade lengths. Visual
representations are also useful for validating and comparing the Figure 4. Alternate strategies for coding temporal order of sac-
expressiveness of different metrics, as well as associating ca- cades in a scaled trace.
nonical visual search strategies with scanpath data.
Though not intended to communicate detailed scanning tenden-
cies, traces can effectively differentiate very horizontal from
2 Methods: Scanpath Visual Representations very vertical scanning tendencies. Figure 5 shows 12 different
trace examples, each with 8 fixations and 7 saccades. This array
Several compact visual representations that summarize scan- of traces is an example of a small multiple, in which minimized
paths are described and compared below. These show scanpath graphics allow the viewer to compare many variables simultane-
fixation density and complexity. Some of the representations ously [Tufte 1983]. In this case, these scanpaths can be com-
show aggregate values associated with revisited locations and pared for relative aspect ratios, horizontal/vertical extent of
aren’t subject to the visual complexity problems associated with scanning, and density of fixations. Saccade crossings in scaled
saccade crossings. The representations can be used to help vali- traces (Figure 5) may provide a rough approximation of scan-
date associations between actual data and prototypical strategies. path complexity, but such conclusions are much less reliable for
longer scanpaths or when comparing scanpaths of different
Scanpaths may also be compressed or simplified, prior to con- lengths.
verting to any of the visual representations. Compression algo-
rithms, defined for specific studies, could merge or delete dupli- While traces provide compact scanpath representations for rapid
cate fixations within close proximity, could remove duplicate comparison, they cannot accurately convey number of fixations,
sub-sequences, or could remove sub-sequences that match in lengths of saccades, or search area. Rather, they only provide
reverse order [Tsai 2001]. relative comparisons of scanpath fixation density and complex-
ity.

204

To construct a radial plot, a scanpath is first broken into its indi-
vidual saccadic segments, much like in Noton and Stark’s origi-
nal scanpath theory [Noton and Stark 1971]. As shown in Table
1, each of these segments can be assigned a length, an absolute
inter-fixation angle, and a relative inter-fixation angle. Rantala
[2008] introduced a radial plot (‘saccade star’) that presents a
distribution of gaze statistics, such as fixation duration, arranged
by angle. The present concept extends this representation to
several other statistics and aggregation methods.

Table 1. Computation of angles in example scanpath.

Saccade ID Length Absolute Relative Angle
Figure 5. Array of scanpath traces, forming a small multiple. (pixels) Angle (o) (o)
Each trace contains 8 fixations and 7 saccades.
1 130 45 --
2.2 Time Expansion
2 55 290 260
Scanpaths and scanpath traces are often tangled with significant
crossing and overlap, making it hard to separate horizontal and 3 52 30 90
vertical tendencies in scanning. By plotting the scanpath x or y-
component against time, ‘time expanded’ representations of 4 206 240 220
scanning tendencies are obtained. The general concept of substi-
tuting time for the horizontal axis of a scanpath has been previ- 5 305 320 100
ously introduced for cases of searching programming code
[Uwano et al. 2006] and search result listings [Raiha et al., 2005, 6 239 110 135
Aula et al. 2005].
7 138 170 40
Separation of the scanpath into vertical and horizontal time ex-
pansions can broaden this concept to scanning whole pages.
Figure 6 shows an 11-fixation scanpath trace that is separated 8 122 320 150
into x and y time expansion graphs. Frequent horizontal shifts
are clearly seen in the x graph, and vertical shifts are apparent in 9 43 75 90
the y graph. Neither of these tendencies is as easily observed in
the original scanpath trace. 10 145 30 310

The time expansion representations are similar to ‘sparklines;’
when plotted as small multiples, they are much like those used
for stock prices and temperature [Tufte 2006]. General move- Absolute angles are measured relative to a global coordinate
ment trends and back-tracking frequency become apparent when system (see Figure 7A), whereas relative angles are measured
several time expansions are compared. relative to the scanpath’s current direction of saccadic move-
ment (see Figure 7B). Scanpath metrics using absolute angles
reveal spatial tendencies relative to element layout on a page,
whereas metrics using relative angles highlight veering from a
set course.

Figure 6. Time expanded x and y representations of scanpath.
Figure 7. Scanpath saccade angles measured counter-clockwise
2.3 Radial Plots from an absolute or global reference frame (A), and a frame that
is relative to the direction of motion (B).
Radial plots are another compact way to represent scanpaths
visually. Radial plots represent scanpath angles and magnitudes The angles are then aggregated into angle ranges, or bins, of a
with bars that radiate from a center point like spokes in a wheel. predetermined resolution, as shown in Table 2. Here, bins are of
width 45o; smaller widths provide greater precision in the plots,

205

at the expense of few samples per bin. Fewer samples per bin, in Example radial plots, based upon the scanpath in Table 1, are
turn, result in less matching because there is less opportunity for shown in Figure 8. Plots based on absolute angles are shown in
different distributions to share the same bins. On the other hand, the left column, and those based upon relative angles are in the
bin sizes that are too large result in less precision for individual right column. The radial histograms, using saccade segment
distributions, but provide more information about how different counts, indicate movement in most of the binned directions, with
distributions are similar. some bias to rightward movement (absolute angles). From the
relative angles, however, there was a small bias in the direction
Each bin is then shown as a bar in the radial plot. Bar angle cor- of 90o movement, left-hand turns, or counter-clockwise scan-
responds to the range of angles in the bin. We typically set each ning. The longest average saccade lengths appeared to be in
bar angle to the angle at the center of the range of angles in each upward and downward directions, although large rightward sac-
bin, but alternate values may be chosen (e.g. the average of the cades were also noted. Notice that only short average saccade
actual angles in the bin). The bar length for each bin may be lengths were made at a relative 0o angle, indicating that the
computed in a number of ways; it may be proportional to: scanpath was indeed quite twisted. The relative angle, summed
lengths were even more sensitive to the twisting than the relative
• Count: The number of saccade segments in the bin (i.e. angle, average saccade length plots.
whose inter-fixation angle lies within the angle range of the
bin). These plots can also be termed radial histograms.
• Average The average length of the saccade segments in the
bin.
• Sum: The sum total length of the saccade segments in the
bin.
• Max (or Min): The maximum (or minimum) saccade seg-
ment length in the bin.

Table 2. Computation of saccade bin counts, with bins of 45o
width and saccade segments from Table 1.

Bin Angles Number of Saccades within Bin
(degrees)
Absolute Relative Angle (o)
Angle (o)

1-45 3 1

46-90 1 2

91-135 1 2

136-180 1 1
Figure 8. Radial plot representations of scanpath from Table 1,
181-225 0 1 based either on absolute or relative angles. The number indi-
cates the maximum value along each axis.
226-270 1 1
3 Results: Strategy Comparison using Repre-
271-315 1 1
sentations
316-360 2 0
3.1 Prototype Strategies

The proposed scanpath representations can be applied to a few
In addition to bin lengths, the mean shift indicates the geometric prototypical scanning strategies, to assess the representations’
or spatial mean of each radial plot. It is represented as a single diagnostic potential for identifying these prototype patterns. A
dot on the radial plot, and is computed from the mean of the x software tool was developed to render the scanpath representa-
and y components of the graph. This represents the overall tions, and to provide a basis for their qualitative and quantitative
movement tendency for the scanpath. The coordinates of this dot comparison. The tool was developed in HTML and JavaScript,
could, optionally, be computed from the weighted means of each and runs in either Firefox or Chrome browsers. It allows cursor-
bin length, where weights are defined by the count (the number based entry of a prototype scanpath trace, which is then repre-
of scanpaths contained) in each bin. Compared to an unweighted sented by each of the plots. Radial plots use default bin widths
mean shift, this weighted mean shift would show greater move- of 20o, but this value is easily modified.
ment in the direction of bins containing greater numbers of
scanpath segments. Table 3 shows nine scanning traces with their compact represen-
tations. These should be considered as examples in a much lar-

206

ger library of prototypical scanning strategies. Sequential traces identical. The absolute angle plots contain data and mean
(the first three columns) are drawn with ink that transitions from shifts that are in the direction of scanning, while the bin an-
light to dark over each temporal sequence. The last two columns gle from the relative angle plots was 0o. Mean shift location
contain an example metric, the number of bins containing data in relative to origin can indicate the direction of scanning.
the radial plots, that could lead to a programmatic solution for
teasing apart the scanning strategies. • Downward, followed by rightward scanning (row 3). X
and Y-time expansions both move downward and to the
By studying these prototypical traces, we can derive a few quali- right. Relative angle radial plots show mostly 0o data, due
tative indicators of various scanning strategies. These indicators to the fact that most saccades are in the same direction as
can, in turn, lead to appropriate metrics to enable more auto- preceding saccades. Absolute angle radial plots are more
mated classification of the scanning strategies. These heuristics balanced between rightward and downward data.
are discussed below, with references to the appropriate row(s)
• Square scanning (rows 4 and 5). Scanpaths that traverse a
from Table 3.
closed loop are indicated by absolute angle radial plots con-
taining data spanning many bins. If scanning is primarily in
• Linear scanning (rows 1 and 2). Whether scanning
one direction, relative angle radial plots show most data in
downward or rightward, both time expansions are linear,
one angular bin. The more looping that is present in a scan-
without reversals, and count and length radial plots are
path, the closer the absolute angle mean shifts are to the
Table 3. Sample prototype scanning strategies, with their compact representations.

207

origin, compared to relative angle mean shifts. In the case ized here, the following are likely to separate scanpath strategies
of a square-like scanning strategy, time expansions show a in classification analyses:
single horizontal and vertical reversal pattern. The radial
plots are primarily single bin (relative angle) and four bins • Number of time expansion reversals. Measures amount
(absolute), whether clockwise (CW) or counter-clockwise of reversal movement in X or Y-components of scanpath
(CCW). Teasing apart CW from CCW direction can be movement. Circles, for example, sweep 1 or 2 reversals,
based upon the primary angle of the mean shift associated depending on starting and ending locations. This measure
with average length, relative angle radial plots. Note that can also be further subdivided into positive and negative
radial plots that are based on count data don’t necessarily directions, to provide further classification precision.
have equal length arms. This is caused by differences in • Number of bins containing data in radial plots. Com-
cursor velocity when inputting the square scanning strat- puted for both absolute and relative angle plots, based on
egy. Here, the absolute angle 90o arm, moving upward, was either counts or scanpath lengths. This metric measures
shorter because it contained fewer fixation counts than the directness and amount of circularity versus squareness in
other arms. This, in turn, was due to faster upward cursor scanpaths. Circles contain many bins, and lines only con-
movement, compared to the other directions. tain a single bin.
• Circular scanning (rows 6 and 7). Both CW and CCW • Angle and Length of mean shift in radial plots. Also
circles showed complex, correlated movement patterns in computed for both absolute and relative angle plots, based
time expansions. Both Relative angle radial plots contain on either counts or scanpath lengths. Measures position
directional data primarily in one or two angular bins. Abso- and distance of the mean shift, relative to 0o origin. Indi-
lute angle radial plots typically contain data in all bins, al- cates primary direction and magnitude of scanpath move-
though the length of individual bins depend on cursor input ment.
velocity (fixation count data) or relative differences in sac-
cade lengths (average length data). Both absolute angle 3.3 Validation: Dataset Analysis
plots contain data that are generally more distributed
among multiple bins than for relative radial plots. If the The process of using the aforementioned heuristics to diagnose
mean shift is located between 0o-180o, there is a tendency scanning tendencies can be illustrated using the results from a
towards counter-clockwise scanning, and clockwise scan- recent study, in which over 120 participants scanned the image
ning otherwise. of a novel iPhone application, shown in Figure 9, in order to
locate the ‘Status’ of the ‘CHI Services Deal’, located in the
• Triangular scanning (row 8). Absolute angle radial plots yellow shaded cell. This relatively simple scanning task resulted
clearly show data in three bins, corresponding to the three in a very rich set of varied scanning strategies.
primary angles, whereas relative plots were primarily one
or two bins, in correspondence with few turns. As for
square scanning, the length of the arms in the absolute an-
gle, fixation count plots depends heavily upon cursor input
speed. Whereas the absolute angle plots contain 3 arms
with data, the relative angle plots contain 2 arms for an
equilateral triangle. These relative angles correspond to 0o,
and 60o (leftward).
• More complex tendencies (row 9). Although not nearly as
complex as ‘real’ scanpath data, a ‘Figure-8’ prototype
scanpath starts to approximate the subtle complexities of
real data. Here, there are more reversals in the X than the Y
time expansion plot, indicating more complex horizontal
than vertical movement. Absolute radial plots show all an-
gular bins are filled, much like circular scanning. Their Figure 9. Emulated iPhone application task image, in which
mean shifts are close to the origin, indicating closed loop participants had to report the status of the CHI Service Deal
scanning. The fact that the relative angle mean shifts were (shaded yellow here).
positive and close to 0o indicates that equal movements
were made rightward and leftward from current scanning 3.3.1 Similar Time Expansions
direction. In even more complex scanpaths, radial plots Table 4 shows a subset of participant scanpaths that contain
based on fixation counts indicate sub-segments with many similar time expansion representations. Due to space constraints,
longer, saccades in a specified direction. Radial plots based we only show the time expansion representations and the scan-
on saccade lengths could indicate a few long saccades in a paths; other figures will show pertinent representations where
specified direction, within a larger set of saccades. there are similarities. Each X-component time expansion indi-
cates an initial leftward movement, followed by a rightward
3.2 Potential Metrics movement. Indeed, each scanpath image validates this pattern.

Based upon the preceding analysis of prototypical strategies, 3.3.2 Similar Radial Plots
metrics can be obtained from the scanpath representations to Similar scanpath sub-sequences can also be identified by radial
help identify prototype strategies within real scanpath data. plots. Table 5 shows absolute and relative angle radial histo-
Many metrics are possible; although these are not further util- grams that contain saccade count data. In each of these scan-
paths, the absolute angle plots had significant bins at 0 o, 180 o,

208

and 270o. Corresponding histograms based on relative angle Table 6. Similar scanpaths, from similar radial histograms.
were dominated by data at 90 o, indicating left turns relative to
current direction. The three scanpaths each travel left, down,
right, up, and down in a similar manner.

Table 4. Similar scanpaths, identified from similar x-component
time expansions.

Table 7. Similar scanpaths, from similar relative angle radial
histograms.

Table 5. Similar scanpaths, identified from absolute and relative
angle radial histograms. Each scanpath was dominated by find-
ing the left table column, scanning down to the bottom row,
locating the ‘Status’ column header, then ending at the desired
cell.
4 Discussion

The scanpath representations introduced in this paper provide
eye tracking and usability specialists with a new tool for identi-
fying scanpaths and scanning strategies. We have described
them, generated examples using prototypical scanpath shapes,
and showed how they can be used to diagnose scanning tenden-
cies.

Scanpaths are fascinating records of human visual attention in
space and time. Scanpaths capture fixation start-time, duration,
and position. Because the samples are time-stamped, they are
ordered and form a sequence.

Each type of representation carries its advantages, much like
providing multiple windows into a dataset. The scaled traces
provide relative comparisons of scanpath fixation density and
complexity. The time expansions help to visually unwind scan-
paths of circular scanning or backtracking. The radial plots and
mean shifts can show individual, aggregated, and multiple scan-
path data. Like heatmaps, they display tendencies within and
across groups and conditions. Unlike heatmaps, however, they
provide directional scanning strategy information. Whether or
Other scanning tendencies were also noted. The two scanpaths not these plots are exposed to the end user, they can serve as a
shown in Table 6 exhibited remarkably similar absolute and basis for metrics that can describe aspects of scanning behavior.
relative histograms. The relative mean shifts were at an angle of It is possible, for example, to count the number of reversals in
about 120 o, indicating lots of leftward turning. Each scanpath the time expansions or the dominant bins in the radial plots.
made a partial CCW circle starting at the center of the table. These metrics could be applied to a set of prototype strategies
and used to train a classification analysis to help tease apart
Radial histograms with many filled angular bins and mean shifts similarities and differences among scanpaths.
near the origin can indicate circular closed loop scanning. This
was observed in Table 7, where both scanpaths made a CCW The compact representations discussed here can be applied to
circle before moving rightward in the table. any path or sequence-related problem with a spatial reference,
not just scanpaths. For example, routing and logistics domains
could benefit from new ways to compare transportation com-
plexities. Radial plots could convey the number, length, and
direction of delivery routes. More abstract spatial analogies,

209

such as comparing imaginary numbers plotted in polar coordi- HORNOF, A.J. AND HALVERSON, T. 2003. Cognitive Strategies
nates to compare filter response, are also possible. and Eye Movements for Searching Hierarchical Computer
Displays, Proc. CHI 2003, ACM Press, 5, 249-256.
We believe that higher-level cognitive strategies can be exposed
by analyzing individual scanpath sequences, then showing ag- JOSEPHSON, S. AND HOLMES, M.E. 2002. Visual Attention to
gregated strategies. As a first step, we have developed visual Repeated Internet Images: Testing the Scanpath Theory on
representations of scanpaths based on data derived from their the World Wide Web. Proc. ETRA 2002, ACM Press, 43-
sequences: saccade lengths, saccade angles, and time expan- 49.
sions. We have also explored how radial plots that aggregate MARSHALL, S.P. 2007. Identifying Cognitive State from Eye
saccades by their inter-fixation angles can be used to character- Metrics. Aviation, Space, and Environmental Medicine,
ize actual scanpaths. 78(5), B165-B185.
Future work will attempt to further understand how compact NOTON, D. AND STARK, L.W. 1971. Scanpaths in Saccadic Eye
scanpath representations can help identify scanning strategies Movements While Viewing and Recognizing Patterns, Vi-
that are exhibited in and across scanpaths. This information, in sion Research, 11, 929-942.
turn, will help eye tracking and usability specialists understand
RAIHA, K., AULA, A., MAJARANTA, P., RANTALA, H., AND KOI-
how people scan images and software user interfaces.
VUNEN, K. 2005. Static Visualization of Temporal Eye-
Tracking Data. Proc. INTERACT 2005, Int. Fed. Info Proc,
946-949.
References RANTALA, H. 2008. Eye2i: Coordinated Multiple Views for
AULA, A., MAJARANTA, P., AND RAIHA, K. 2005. Eye-tracking Gaze Data. Proc. ETRA 2008, ACM Press, 147.
Reveals the Personal Styles for Search Result Evaluation, SANTELLA, A., AND DECARLO, D. 2004. Robust Clustering of
Proc. INTERACT 2005, Int. Fed. Info Proc, 1058-1061. Eye Movement Recordings for Quantification of Visual In-
FEUSNER, M., AND LUKOFF, B. 2008. Testing for Statistically terest. Proc. ETRA 2004, ACM Press, 27-34.
Significant Differences Between Groups of Scan Patterns. TSAI, C. 2001. A C-program for Scanpath Compression,
Proc. ETRA 2008, ACM Press, 43-46. http://research.chtsai.org/papers/scanpath-
GOLDBERG, J.H., AND KOTVAL, X.P. 1999. Computer Interface compression.html.
Evaluation using Eye Movements: Methods and Constructs. TUFTE, E. 1983. The Visual Display of Quantitative Information.
Int. J. Industrial Ergonomics, 14, 631-645. Graphics Press, Cheshire, CT., 170-175.
GOLDBERG, J.H., AND KOTVAL, X.P. 1998. Eye Movement-Based TUFTE, E. 2006. Beautiful Evidence. Graphics Press, Cheshire,
Evaluation of the Computer Interface, in Kumar, J. (Ed.), CT.
Advances in Occupational Ergonomics and Safety, IOS
Press, 529-532. UWANO, H., NAKUMURA, M., MONDEN, A., AND MATSUMOTO, K.
2006. Analyzing Individual Performance of Source Code
HEMBROOKE, H., FEUSNER, M., AND GAY, G. 2006. Averaging Review Using Reviewers’ Eye Movement. Proc. ETRA
Scan Patterns and What They Can Tell Us. Proc. ETRA 2006, ACM Press, 133-140.
2006, ACM Press, 41.
WEST, J.M., HAAKE, A.R., ROZANSKI, E.P., AND KARN, K.S.
HEMINGHOUS, J., AND DUCHOWSKI, A.T. 2008. iComp: A Tool 2006. eyePatterns: Software for Identifying Patterns and
for Scanpath Visualization and Comparison. PROC APGV Similarities Across Fixation Sequences. Proc. ETRA 2006,
2008, Symposium on Applied Perception in Graphics and ACM Press, 149-154.
Visualization, ACM Press, 152.
YARBUS, A.L. 1967. Eye Movements and Vision. New York,
Plenum Press.

210

Goldberg Visual Scanpath Representation

More Related Content

What's hot (15)

Viewers also liked (20)

Similar to Goldberg Visual Scanpath Representation (20)

More from Kalle (20)

Goldberg Visual Scanpath Representation