Gaze Transition Entropy, KRZYSZTOF KREJTZ, ANDREW DUCHOWSKI

1
Gaze Transition Entropy

KRZYSZTOF KREJTZ, National Information Processing Institute, Warsaw, Poland and University of
Social Sciences and Humanities, Warsaw, Poland
ANDREW DUCHOWSKI, Clemson University, Clemson, SC, USA
TOMASZ SZMIDT, Information Processing Institute, Warsaw, Poland
IZABELA KREJTZ, University of Social Sciences and Humanities, Warsaw, Poland
FERNANDO GONZALEZ PERILLI, Center for Basic research in Psychology, UdealR, Uruguay and
Department of Basic, Evolutionary and Educational Psychology, Universitat Autonoma de Barcelona,
Spain
ANA PIRES, Center for Basic research in Psychology, UdealR, Uruguay and Department of Basic,
Evolutionary and Educational Psychology, Universitat Autonoma de Barcelona, Spain
ANNA VILARO, Department of Basic, Evolutionary and Educational Psychology and Center for
research in Ambient Intelligence & Accessibility in Catalonia (CAIAC), Universitat Autonoma de
Barcelona, Spain
NATAlIA VILLALOBOS, FULVIO CAPURSO, ROSSIO GONI, EMILIANO FERNANDEZ, Center for
Basic research in Psychology, UdealR, Uruguay
The paper introduces a two-step method of quantifying eye movement transitions between Areas Of Interests (AOIs). First,
individuals gaze switching patterns, represented by fixated AOI sequences, are modeled as Markov chains. Second, Shannons
entropy coefficient of the fit Markov model is computed to quantify the complexity of individual switching patterns. To determine
the overall distribution of attention over AOIs, the entropy coefficient of individuals stationary distribution of fixations is
calculated.
The novelty of the method is that it captures the variability of individual differences in eye movement characteristics, which
are then summarized statistically. The method is demonstrated on gaze data collected during free viewing of classical art
paintings. Shannons coefficient derived from individual transition matrices is significantly related to participants individual
differences as well as to their aesthetic experience of art pieces.
Categories and Subject Descriptors: J.4 [Computer Applications]: Social and Behavioural SciencesPsychology.
General Terms: Human Factors
Additional Key Words and Phrases: eye movement transitions, entropy
Contact authors addresses: K. Krejtz, Laboratory of Interactive Technologies, Institute of Information Processing, 00-680 War-
saw, Poland, and Department of Psychology, University of Social Sci. and Humanities, 03-815 Warsaw, Poland; email: kkre-
jtz@opi.org.pl.
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 for profit or commercial advantage and that copies show this notice on the first page
or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to
lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permissions may be
requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481,
or permissions@acm.org.
c 2014 ACM 1544-3558/2014/05-ART1 $15.00
DOI 10.1145/0000000.0000000 http://doi.acm.org/10.1145/0000000.0000000
ACM Transactions on Applied Perception, Vol. 2, No. 3, Article 1, Publication date: May 2014.
1:2 K. Krejtz et al.
1. INTRODUCTION
There is a pressing need for quantitative comparison of eye movement metrics [Duchowski et al. 2010].
There are two leading eye movement metrics for which quantitative difference methods have been de-
veloped. The first are scanpaths, represented by an ordered sequence of fixations, for which vector and
string-based editing approaches have been developed to compute similarity differences [de Bruin et al.
2013; Jarodzka et al. 2010]. The second are heatmaps, represented (generally) by Gaussian Mixture
Models (GMMs) indicating frequency (or probability) of fixation localization. Lacking sequential order
information, heatmap-based metrics generally involve a measure of overlap between two GMMs, in-
dicating similarity of fixated regions of an image, but not their order of visitation [Grindinger et al.
2010a; Grindinger et al. 2010b].
Transition matrices are a third type of sequential gaze pattern analysis but for which quantitative
similarity measures are scarce. In this paper, we present a method for comparing transition matrices
and distribution of fixations between AOIs. The utility of quantitative comparisons between gaze pat-
tern representations usually lies in the techniques ability to produce tests of statistical significance.
We extend Krejtz et al.s [2014] paper by addressing the following important details omitted previ-
ously:
(1) handling of matrix sparsity, e.g., when no observed fixation transition from a given source AOI are
recorded,
(2) entropy normalization,
(3) computation of the stationary distribution estimate via eigenvalue analysis, and
(4) effects of content-independent AOI grid granularity variation.
2. BACKGROUND
Ponsoda et al. [1995] introduced the probability vector and transition matrix to categorize saccade
directions. Interestingly, their matrices were based on transitions between eight cardinal (compass)
saccade directions (e.g., N, NE, SE, S, etc.) instead of from and to Areas Of Interest, or AOIs, as is
more commonly done today. Although they used Z and 2 statistics to compare matrices, they did not
consider the sequence of saccade directions to be a Markov chain, nor did they employ any Markov
chain properties to obtain conclusions about saccade directions. In this paper we focus on saccadic
transitions between AOIs instead of cardinal directions and model them as Markov chains. A Markov
chain is a stochastic system undergoing transitions between states, in which the next state depends
only on the present one, and does not depend on the past. Such random processes have found many
applications in social, technical and life sciences.
Goldberg and Kotval [1999] discussed transition matrices between AOIs as one of several eye move-
ment comparison metrics in usability evaluation, including: number of saccades, saccade length and
duration, saccadic amplitude, convex hull area, spatial density, number of fixations, fixation durations,
and a fixation/saccade ratio. They noted the difference between content-dependent AOI assignment to
specific screen objects and content-independent AOI assignment where the screen is simply divided
into a uniform grid. To compare matrices, they suggested transition matrix densitythe number of
non-zero matrix cells divided by the total number of cellsas an indicator of matrix sparsity. In this
paper we introduce transition matrix entropy as a means of matrix comparison. Entropy estimation
is a histogram-based method which in effect takes into account the width of the histogram bins (AOI
size), an aspect that density alone may miss if it only considers the total number of cells.
Liechty et al. [2003] used Hidden Markov Models (HMMs) to differentiate between local and global
covert visual attention states from eye movement data. Instead of applying the Markov model to tran-
Gaze Transition Entropy 1:3
sition matrices as we do in this paper, their use of HMMs was aimed at distinguishing fixations, akin
to Velichkovsky et al. [2005], who suggested classifying fixations as ambient or focal.
Bednarik et al. [2005] constructed transition matrices to describe visual switching behavior among
predefined regions, but found no correlation between task performance and this high-level data. Fis-
cher and Peinsipp-Byma [2007] combined transition matrices and descriptive statistics extracted from
a uniform grid sampling of the stimulus to compare individuals perception of the stimulus. Neither of
these efforts compared transition matrices, however.
Hwang et al. [2011] introduce transitional semantic guidance computation to evaluate fixation tran-
sitions. Instead of constructing a transition matrix from AOIs or a grid, they considered transitions
from locations within a generated saliency map of a given scene. Semantic saliency of the next fix-
ation target determined the guidance score of that gaze transition. This method can be considered
as something of a hybrid between transition matrix construction and scanpath comparison where
transition matrices are replaced by semantic maps. One conceivable disadvantage to this approach
is comparison between scenes since saliency maps generated for each are likely to differ. Constructing
content-independent transition matrices would overcome this problem.
Finally, the most recent example of work similar to ours is that of Vandeberg et al. [2013]. They
analyze eye movement transitions with a multilevel Markov modeling approach. However, when com-
paring transition matrices, they compare matrix entries (transition probabilities) directly, whereas we
use the entropy of individual transition matrices as a means of transition matrix comparison (advan-
tages of our approach are made clear in the following section).
Our approach consists of two entropy measurements: transition entropy (calculated for individual
subjects transition matrices) and stationary entropy (calculated for individual subjects stationary
distributions) and allows for achieving straightforward measures of predictability (reflexiveness) in
AOI transitions and overall distribution of eye movements over stimuli. Moreover, these two measures
can be easily employed for comparisons with common statistical procedures.
Although the usefulness of entropy for analyzing eye movements has been previously demonstrated,
the manner of usage of this measure differs significantly [Althoff 1998; Acarturk and Habel 2012; Shic
et al. 2008; Kruizinga et al. 2006]. For instance, Acarturk and Habel [2012] examined visual compre-
hension of graphs in multimedia environments and calculated entropy for summary transition matri-
ces where the cells values represent the number of transitions between certain AOIs. This method,
however, did not allow for discovering significant differences or other relations via statistical tests.
Similarly, Shic et al. [2008] used entropy measures to find differences in the visual exploration of faces
by autistic and typically developing children. They, however, did not distinguish between transition
and stationary entropy estimates.
The present paper describes the method for achieving the two measures of entropy and their prop-
erties, and shows their usefulness when applied to an example of a number of empirical examples.
3. MARKOV PROPERTY OF GAZE TRANSITIONS

A common practice while analyzing saccadic transitions between AOIs is to calculate transition ma-
trices for each group of subjects under given experimental conditions. The value in row i and column j
denotes empirical probability that the next fixation would be to the j th AOI if it was presently in the
ith AOI.
In the present paper, we apply a different approach based on transition matrices calculated for every
subject, not groups of them. The main advantage is that we can analyze individual switching patterns
between AOIs, which can vary significantly between subjects. The limitation lies in the number of
AOIs defined for a stimulus. If there are too many, some rows of the transition matrix can be equal
to zero due to an insufficient amount of data. This problem is addressed in the implementation notes,
below.
Consider one subject and let Xt be a random variable taking values in the set of AOIs S = {1, . . . , s}
(the state space) and denoting where this subject focused their attention during the tth fixation (t =
1, . . . , n). Analyses of transition matrices require ensuring that the next AOI depends only on the
current one and that it is not affected by the past. Precisely, we need
P(Xt+1 = xt+1 |Xt = xt , . . . , X0 = x0 ) = P(Xt+1 = xt+1 |Xt = xt ),
where xt S . This condition is called the Markov property and a stochastic process satisfying this
is called a first-order Markov chain (or just a Markov chain). When the future state depends on both
the present state and the one before, we can define a second-order Markov chain, and so on. The
process describing transitions of the subjects visual attention that follows the Markov property is
fully determined by the initial state X0 = x0 and transition matrix P = (pij )ss , where pij = P(Xt+1 =
j|Xt = i), t = 0, . . . , n 1 is the probability of changing the fixations gaze position from the state i
to j. We shall validate the quality of fitting a temporally homogeneous Markov chain (i.e., switching
probabilities that do not depend on t) to the observed individual switching patterns.
3.1 Testing of the Markov Property

To our knowledge, the Markov property has not yet been tested in modeling AOI sequences. Let us
describe the test procedure presented by Besag and Mondal [2013]. Suppose we are given the Markov
chain with transition matrix P and initial state x0 . By the chain rule, the probability of obtaining the
observed sequence x = (x0 , . . . , xn ) with frequency counts nij of transitions from state i to j is equal to
the product
Y n
pijij , (1)
i,jS
which is called the likelihood of the given sequence. The common log-likelihood ratio test statistic takes
the form of
nijk
Q n+jk
i,j,k n+j+
u = 2 log Q nijk . (2)
nijk
i,j,k nij+
where nijk is the frequency count of the AOI triple (i, j, k) in the sequence x and + denotes the sum-
mation over all AOIs. The numerator in the expression above is the likelihood of the data when we
assume the null hypothesis that the sequence follows a Markov chain of order 1 and the denomina-
tor is the likelihood of the alternativea Markov chain of order two. A more complex model fits the
data better, so the test statistic is positive and its high value suggests rejecting the assumption of the
Markov property.
Notice that expression (2) can be transformed to the form
X Oijk
u=2 Oijk log , (3)
Eijk
i,j,kS
+jk n
where Oijk = nijk is the observed frequency and Eijk = nij+ n+j+ is the expected frequency of a triple
(i, j, k) under the null hypothesis assumption. The test statistic is proportional to the Kullback-Leibler
divergence [Kullback and Leibler 1951] of the observed distribution of triples from the theoretical when
the Markov property holds.
When the null hypothesis is valid, the test statistic has an asymptotic distribution 2 with s(s 1)2
degrees of freedom. AOI sequences are often short, causing the distribution of the test statistic to be far
from the asymptotic limit. Besag and Mondal [2013] proposed an exact test, based on random sampling
from the space of all sequences starting from the same state x0 with transition counts nij to be the
same as the observed sequence x. The p-value of the test is then calculated as the share of sequences
with a higher value of the test statistic than the observed one.
During analysis the theoretical probabilities in the expressions above are replaced with their maxi-
mum likelihood estimators, as assumed by Ciuperca and Girardin [2005],
nij ni
pij = , i = , i, j S. (4)
ni n
4. ENTROPY OF AOI SEQUENCE

Suppose we are given a sequence x = (x0 , . . . , xn ) modeled as a Markov chain which describes the
subjects AOI switching pattern, with transition probabilities pij that are stationary, i where i, j S.
Estimation of i is discussed in implementation notes below. The complexity of the process can then be
measured by Shannons entropy [Ekroot and Cover 1993; Ciuperca and Girardin 2005]:
X X
Ht = i pij log2 pij . (5)
iS jS
The highest possible entropy equal to log2 s is reached when the distribution of transitions is uni-
form for each AOI (which also means that subsequent transitions are independent of each other). The
minimal entropy of 0 describes a fully deterministic Markov chain, i.e., one in which all transition
probabilities are either 0 or 1. Thus, the higher the entropy, the more randomness there is in a sub-
jects transitions, the more complex the sequence of AOIs and the more exploratory the character of
visual attention.
Stated another way, entropy refers to the expected surprise of a given gaze transition. Minimum
entropy of 0 suggests no expected surprise, meaning that a gaze transition is always expected to the
same j th AOI. Maximum entropy, on the other hand, suggests maximum suprise, since transition from
source AOI to any destination AOI is equally likely, and hence whichever occurs expects in maximum
surprise. More formally, the term pij log2 pij in (5) is the transitions contribution to system entropy,
modeled by its probability multiplied by its surprisal (or information content) [Hume and Mailhot
2013].
We are also interested in the entropy of the stationary distribution itself:
X
Hs = i log2 i . (6)
iS
A higher value of stationary entropy means that the subject distributes their visual attention more
equally between AOIs. A lower value is obtained when fixations tend to be concentrated on certain
AOIs, possibly because they appear to be more interesting to the viewer.
One can expect that the long-term distribution of fixations over AOIs should be more uniform than
the transitions distributions, i.e. the stationary entropy is greater than the transition one. The for-
mal proof is as follows. For all states the KullbackLeibler divergence of the stationary distribution
(1 , . . . , s ) from the transition distribution (pi1 , . . . , pis ) is non-negative, which is a direct consequence
of the Jensens inequality1

X pij
pij log2 0, i = 1, . . . , s. (7)
j
j
We multiply both sides by stationary probabilities i and take the sum over all states, so
X X X X
i pij log2 pij i pij log2 j 0. (8)
i j i j
The first double sum is equal to Ht . In the other one we rearrange the summation order
X X
Ht log2 j i pij 0. (9)
j i
P
Let us notice that i i pij is by definition equal to j , thus the inequality
Hs Ht (10)
is proved.
5. IMPLEMENTATION
Transition matrix P can be computed in R [R Development Core Team 2011] for each of the AOIs
defined atop the stimulus image. Matrix elements pij are set to the number of transitions from ith
source AOI to j th destination AOI for each participant and then the matrix is normalized relative to
each source AOI (i.e., per row),
nij
pij = P , i, j S, (11)
j nij
such that pij represents the estimated probability of transitioning from ith AOI to any j th AOI given
ith AOI as the starting point.
5.1 Entropy Normalization
In practice, it is possible that no transitions from the ith AOI are observed. This leads to
Pa zero matrix
row sum in (11) and division by zero. When this occurs, i.e., the matrix row sum is zero ( i pij = 0) , we
handle this happenstance by setting each of the row entries to their uniform transition distribution,
namely pij = 1/s, assuming an equally likely probability of transitioning to any other AOI given this ith
source AOI. Doing so for all pij would lead to a uniform matrix with maximum entropy equal to log2 s
bits per transition. Indeed, maximum entropy is used to normalize the empirical entropy obtained from
each transition matrix.
Empirical entropy is computed by the maximum likelihood method as implemented by Rs publicly
available entropy module [Hausser and Strimmer 2009].
5.2 Statistical Comparison
To facilitate statistical comparison of mean entropies per experimental condition, Ht is computed per
individual participant and per condition and normalized (divided by log2 s). This results in a table of
entropies (each entropy computed from an individuals transition matrix) for each of experimental con-
ditions and each of the participants. Analysis of variance (ANOVA) is then used to test for differences
in mean (normalized) entropy per condition.
1 Typically,
the KL divergence takes the natural logarithm, here we use the base of 2 to measure the entropy in bits, in fact units
do not matter.
5.3 Stationary Entropy Implementation via Eigen Analysis

Stationary entropy Hs and stationary distribution are estimated through eigenvalue analysis. Tran-
sition matrix P represents the likelihood of the state of the system next period. The likelihood of the
state in two periods is given by P2 = PP. Similarly, the likelihood of the state in the long run is given
by
Pn = Pn .
If the transition matirx P is regular, then, over a sufficient number of periods n, the system approaches
a fixed state and transition probabilities will converge to , i.e.,
xPn , x.
The steady-state (stationary) probability vector can be solved for directly if the following holds:
P = .
This equation is satisfied if the steady-state vector is an eigenvector of P corresponding to eigenvalue
= 1.
Given transition matrix P, Rs eigen module returns both of its eigenvalues and eigenvectors. The
first eigenvector e corresponds to eigenvalue = 1, and its normalization gives the stationary distri-
bution, = e/kek. Stationary entropy Hs is then computed via
1 X
Hs = i log2 i , (12)
log2 s i
in agreement with (6) save for the base 2 logarithm which is used to yield entropy in units of bits per
transition and the normalization factor which yields normalized stationary entropy.
5.4 Example Computation
An example calculation for a subject looking at three Areas Of Interest yielded the following transition
matrix P and vector of stationary probabilities :

0.52 0.39 0.10
P = 0.32 0.35 0.32 , = (0.32, 0.31, 0.37) , (13)
0.15 0.21 0.64
which have entropies Ht = 0.94 and Hs = 0.99, respectively (see example below).
6. EXAMPLE APPLICATIONS OF TRANSITION ENTROPY ANALYSIS

To demonstrate the utility of our entropy-based metrics for analysis of eye movement transitions, we
apply it to two practical examples of viewing artwork. In the first example, artwork from three periods
is viewed by observers and their gaze transitions are evaluated in relation to their personal curiosity
and interest in the artwork. In the second, we compare viewers gaze transitions over a stylized ren-
dering of La Jaconde in relation to their self-reported recognition of the paintings representation. In
both examples, we consider the effects of (content-independent) AOI grid granularity. The first example
splits the image vertically, the second tesselates the image into rectangular grid cells.
6.1 Viewing Artwork in Three Classical Styles
The first example compared gaze transitions from viewing paintings from three classical peiord: Im-
pressionist, Renaissance, and Bauhaus. Specifically, three paintings (digital reproductions of oil paint-
ings) were used as stimulus: Jour de pluie a Paris by Gustave Caillebotte (Impressionism), The
(a) Renaissance2 AOIs (b) Renaissance3 AOIs
Fig. 1. An example of Renaissance stimulus with 2 and 3 AOIs marked as used in the experiment.
Tempest by Giorgione (Renaissance), and Composition VII by Vassily Kandinsky (Bauhaus). We an-
alyze data from a study conducted by Biele et al. [2013], wherein gaze data was analyzed in terms of
ambient or focal fixations, as per Velichkovsky et al. [2005]. Our entropy-based transition analysis is
newly applied to their data.
6.1.1 Experimental Design and Procedure. The free viewing experiment consisted of a presentation
of the paintings (Impressionist, Renaissance, or Bauhaus), with each painting presented for 30 seconds
on the computer screen. An example of the artwork, with 2 and 3 AOIs defined for the analysis (marked
with green lines) is shown in Figure 1 2 . AOIs on the rest of stimuli were marked in a similar way.
The experimental procedure consisted of three stages. First, all participants filled out the Curiosity
and Exploration Inventory (CEI-II) [Kashdan et al. 2009]. Second, participants were randomly as-
signed to one of the experimental conditions: positive or neutral mood. Mood induction followed the
procedure proposed by Velten [1968]. Participants in both conditions were presented with 20 positive
or emotionally neutral sentences. Each sentence was presented for 12 seconds on the computer screen.
The participants task was to read all of the sentences and for each of them to think of relevant situa-
tions in their own life. During reading, classical music, emotionally consistent with the sentences, was
played [Richell and Anderson 2004; Robinson et al. 2012]. Third, participants were asked to view art
paintings randomly displayed on the computer screen. Their task was . . . to examine artwork as they
would look at them in an art album. After each artwork presentation they evaluated its attractiveness
on a 5-point Likert-type scale.
The experiment followed a factorial experimental design. The main independent variables were ex-
perimental condition (positive vs. neutral mood), curiosity factor based on median split of the CEI-II
scores, and subjective attractiveness of the artwork (low vs. highanswers from 1 to 3 were coded as
low and answers from 4 to 5 as high attractiveness). The main dependent variables were transition
(Ht ) and stationary (Hs ) entropies.
2 The picture of the artwork was retrieved from http://en.wikipedia.org/wiki/File:Giorgione_019.jpg. Source/photogra-

pher: Florian Heine Das erste Mal. Wie Neues in die Kunst kam, Munchen: Bucher Verlag 2007, ISBN 978-3-7658-1511-9
(a) Ht = 0.89, Hs = 1.00 (b) Ht = 0.69, Hs = 0.98 (c) Ht = 0.71, Hs = 0.78
1 0.52 0.39 0.10 1 0.71 0.21 0.08 1 0.33 0.53 0.13

Source AOI (from)
Source AOI (from)
Source AOI (from)

2 0.32 0.35 0.32 2 0.18 0.61 0.21 2 0.12 0.75 0.12
3 0.15 0.21 0.64 3 0.03 0.19 0.77 3 0.05 0.47 0.47

1
3
Destination AOI (to) Destination AOI (to) Destination AOI (to)
(d) Ht = 0.89 (e) Ht = 0.69 (f) Ht = 0.71
Fig. 2. Scanpaths and transition matrices with low and high entropy values.
6.1.2 Participants. Forty nine university students (31 male and 18 female, aged M = 24.47, SD =
6.53) took part in the experiment after signing a consent form. Data from six subjects were excluded
from the analysis due to eye tracking system calibration problems resulting in a low tracking ratio or
due to insufficient amount of fixations or zero rows in the transition matrix.
6.1.3 Apparatus. All stimuli were presented on a computer monitor (1680 1050 resolution; 22
inch LCD, 60 Hz refresh rate) connected to a standard PC computer. Eye movements were recorded
at 120 Hz with an SMI eye tracking system. SMIs BeGaze software was used for fixation and saccade
detection with a dispersion-based event detection algorithm. The dispersion was 100 px, with minimum
fixation duration set to 80 ms. When analyzing visual exploration patterns it is frequently accepted to
use fixations shorter than 100 ms [Velichkovsky et al. 2005]. Data analyses as well as calculation of
transition matrices and H coefficients were conducted with R [R Development Core Team 2011].
6.2 Results
Analysis of the eye movement data focuses on comparison of subjects transition and stationary en-
tropies, in relation to their self-reported characterization of curiosity and their evaluation of the at-
tractiveness of the freely viewed artwork.
Table I. P-values of the exact (asymptotic) Markov chain

goodness-of-fit test for individual eye movements
transitions over 2 and 3 AOIs.
No. AOIs p-value Impress Renaiss Bauhaus
2 AOIs p 0.05 1 (2) 5 (6) 3 (2)
0.05 < p 0.1 4 (4) 2 (5) 0 (3)
p > 0.1 38 (37) 36 (32) 40 (38)
3 AOIs p 0.05 3 (0) 5 (5) 4 (1)
0.05 < p 0.1 3 (2) 8 (6) 1 (0)
p > 0.1 37 (41) 36 (32) 38 (42)
6.2.1 Results of Markov Chain Fitting. For each stimulus and number of AOIs we checked the
resultant AOI sequences whether modeling with a Markov chain was valid. Results of the exact and
asymptotic test3 described in Section 3 are presented in Table I.
One can see that the assumed model fits the data reasonably well. Only a few cases occurred for
which the hypothesis of Markov property should be rejected (at a significance level of 0.05). This is
confirmed by the test for multiple sequencesfor all three stimuli the probability was close to 1, giving
no reason to doubt that AOI sequences are independent and come from Markov chains with individual
transition matrices.
6.2.2 Transition and Stationary Entropies. Splitting into 3 Areas of Interests is considered in this
subsection. Scanpaths of three subjects with strongly differing entropy values are depicted in Figure 2.
The first two images present AOI sequences with a similar stationary entropy Hs and different value
of Ht . In Figure 2(a) the subject frequently switched between male and female figures resulting in high
transition entropy. The low value of Ht indicates more careful viewing of AOIs, which is reflected in
longer subsequences of fixations in the same area (Figure 2(b)).
The last two images present scanpaths with similar switching complexity, but with highly different
stationary AOI distributions. In Figure 2(c) the subject was mostly interested in the sky, rather than
the woman nursing her child and paid little interest to the man. In Figure 2(b) the subjects visual
attention was split more evenly between AOIs, yielding a higher value of Hs .
Transition matrices of three selected scanpaths are depicted in Figures 2(d)2(f). More uniform tran-
sition probabilities in subsequent rows result in a higher value of Ht .
6.2.3 Analysis of Observed Entropy. Figure 3 shows the exemplary distribution of Ht and Hs among
subjects for Renaissance artwork. The distributions of Ht for all stimuli are Gaussian in shape. Shapiro-
Wilk tests of normality conducted separately for each stimuli divided into 2 and 3 AOIs produced in-
significant results, whereas tests of normality performed for Hs showed that all distributions deviated
from a normal shape. Therefore, we recommend considering nonparametric statistical tests or proce-
dures for dealing with stationary entropy data as a dependent variable.
Recall that in the case of a uniform distribution over the subset {1, . . . , s}, the entropy is equal to
log s. This leads to the prediction that significant differences in mean values of the entropy should be
observed between stimuli with 2 and 3 AOIs. Moreover, we suspected significant differences between
the stimuli, reflecting various art styles and content. In order to test these hypotheses a within-subjects
2 3 analysis of variance was performed. In this analysis the number of AOIs (2 vs. 3) and stimuli
(Impressionism vs. Renaissance vs. Bauhaus) were tested as within-subject factors and the transition
entropy values as the dependent variable. In line with the hypothesis, analysis revealed a statistically
significant main effect of number of AOIs, F (1, 41) = 808.75, p < 0.001, p2 = 0.667. Mean transition
3 We used testing procedure in R written by Besag and Mondal [2013].

Renaissance 2 AOI Renaissance 3 AOI
6 6
Count
Count
4 4
2 2
0 0
0.25 0.50 0.75 1.00 1.25 0.25 0.50 0.75 1.00 1.25
Transition entropy Transition entropy
Renaissance 2 AOI Renaissance 3 AOI
30 30
Count
Count
20 20
10 10
0 0
0.4 0.6 0.8 1.0 1.2 0.4 0.6 0.8 1.0 1.2
Stationary entropy Stationary entropy
Fig. 3. Exemplary distributions of transition and stationary entropies for Renaissance artwork with 2 and 3 AOIs.
entropy for 3 AOIs was significantly higher (M = 0.73, SE = 0.01) than the mean transition entropy
value for 2 AOIs (M = 0.47, SE = 0.01).
Results also showed a significant main effect of stimuli, F (2, 82) = 63.02, p < 0.001, p2 = 0.281. The
mean transition entropy values were the highest for the Renaissance stimulus (M = 0.61, SE = 0.01),
and similarly lower for the Impressionism (M = 0.55, SE = 0.02) and Bauhaus (M = 0.55, SE = 0.01)
stimuli. Pairwise comparisons with Bonferroni correction showed that only the differences between
Renaissance and the other stimuli were significant. The analysis also showed the interaction effect to
be significant, F (2, 82) = 6.97, p < 0.001, p2 = 0.025 (see Figure 4).
6.2.4 Participants Individual Differences of the Entropies. One of the advantages of the proposed
approach is the possibility of a statistical description of the dependence of eye movement transitions on
individual differences. To demonstrate this, in the present experiment we may consider the hypothe-
ses that individuals transitions differ depending on their individual curiosity, mood and stimuli. We
expected that highly curious people would exhibit more focused eye movement patterns (i.e., concen-
trating longer on the same AOI = low transition entropy values), while less curious people would
move their visual attention in a more chaotic way (switching frequently between different AOIs =
high transition entropy values). We also hypothesized based on prior research, e.g., that of Biele et al.
[2013], that participants eye movement patterns would depend on their mood.
In order to test the above hypotheses a 2 2 3 analysis of variance was conducted with Ht as the
dependent variable. The CEI-II score median-split with high vs. low scores, was tested as a between-
subject factor as well as the experimental condition (induced mood: neutral vs. positive), and stimuli
(Bauhaus vs. Impressionism vs. Renaissance) as a within-subject factor. The analysis of variance was
run for entropies calculated for 2 AOIs.
The analysis revealed significant interaction between curiosity and stimuli, F (2, 76) = 3.46, p <
0.05, p2 = 0.035 (see Figure 5). Participants with high curiosity scores had lower Ht values (M =
0.42, SE = 0.02) than subjects with low curiosity scores (M = 0.47, SE = 0.02) while looking at the Bauhaus
Normalized Entropy vs. Task Type Normalized Entropy vs. Task Type
(2 AOIs) (3 AOIs)
1.00 1.00
Mean Normalized Entropy (bits/transition with SE)

0.75 0.75
0.50 0.50
0.25 0.25
Entropy type Entropy type

Stationary Stationary
Transition Transition
0.00 0.00
Bauhaus Impressionism Renaissance Bauhaus Impressionism Renaissance

Artwork type Artwork type
Fig. 4. Differences in transition and stationary entropies vs. number of AOIs and stimulus. Error bars represent 1SE.
CEI score
Low
High
0.6
Transition entropy
0.4
0.2
0.0
Bauhaus Impressionism Renaissance

Stimuli
Fig. 5. Differences in Ht depending on curiosity and stimuli. Error bars represent 1SE.
style painting (p = 0.07). For Impressionism and Renaissance paintings the difference in Ht between
participants with high and low CEI scores was not significant. The analysis showed also the main
effect of stimuli, F (2, 76) = 36.61, p < 0.001, 2 = 0.279. The highest transition entropy score was ob-
tained for Renaissance painting (M = 0.54, SE = 0.01) compared to Bauhaus (M = 0.45, SE = 0.01) and
Impressionism painting (M = 0.43, SE = 0.01). All effects involving condition factor were insignificant.
The above results partially supported our prediction that highly curious participants produce lower
transition entropy levels while viewing the artwork. This can mean that their pattern of visual scan-
ning is more locally oriented. They carefully scan the first AOI devoting many fixations to it before
moving to the second one. Thus it is important to verify whether curious participants distribute their
attention over the whole picture or just focus on a selected area. Considering this scenario, we con-
ducted a similar 2 2 3 analysis of variance with the same factors as above but with stationary
entropy (Hs ) as the dependent variable. Hs describes the final distribution of gaze transitions between
all AOIs. Due to highly skewed distributions of Ht values before running the analysis we followed
the aligned rank transform for nonparametric factorial data analysis [Wobbrock et al. 2011]. Contrary
to the given supposition, all statistical effects of the analysis lacked significance: the main effects of
experimental condition, curiosity, and interaction were all below statistical significance (F < 1).
6.2.5 Relation Between Eye Movement Transitions and Attractiveness of the Artwork. The differ-
ences between subjects eye movement transitions while visually investigating the artwork may also
coincide by their evaluation of the artwork. One may expect that people who like the image more will
also scan it with their eyes more carefully, i.e., in a more predictable way, resulting in lower Ht values.
To test this hypothesis we performed a 2 2 analysis of variance, where between-subject factors were
personal evaluation of attractiveness of the stimuli (low vs. high) and experimental condition (positive
vs. neutral mood). Transition entropy served as the dependent variable.
The main effect of condition was not significant F < 1. The main effect of subjective evaluation of
artwork reached statistical tendency level F (1, 69) = 3.59, p = 0.06, p2 = 0.049. As expected, transition
entropy values for participants who liked the artwork tended to be lower (M = 0.46, SE = 0.01) than
for those who liked the artwork less (M = 0.49, SE = 0.01). Results revealed a significant interaction
effect between artwork evaluation and experimental condition, F (1, 69) = 4.92, p < 0.05, p2 = 0.067 (see
Figure 6). The following pairwise comparisons with Bonferroni correction showed that in the positive
condition those participants who highly valued the art pieces produced significantly (p < 0.01) lower
transition entropy values (M = 0.44, SE = 0.02) compared to those whose evaluation was lower (M =
0.50, SE = 0.03). In the neutral condition, the difference between those who evaluated the artwork
highly or moderately was not significant (p > 0.1).
The above analysis of transition entropy was followed by a similar analysis of variance with station-
ary entropy as the dependent variable. Due to highly skewed distributions of stationary entropy, its
values were preprocessed following the aligned rank transform for nonparametric factorial data anal-
ysis [Wobbrock et al. 2011]. The analysis revealed a statistically significant main effect of subjective
attractiveness, F (2, 69) = 4.68, p < 0.05, p2 = 0.064. This main effect shows that mean rank of stationary
entropy was significantly higher (M = 35, SE = 2.42) among those who evaluated artworks highly than
among those who evaluated artworks relatively lower (M = 27, SE = 2.12). Neither main effect of exper-
imental condition (F < 1) nor the interaction term (F (1, 69) = 1.06, p > 0.1) were statistically significant.
6.3 Viewing La Jaconde

The second example compared gaze transitions from viewing a stylized representation of La Jaconde
(see Fig. 7). We analyze data from a study conducted by Perilli et al. [2014], wherein gaze data was
analyzed in terms of traditional eye movement metrics. Our entropy-based transition analysis is newly
applied to their data.
6.3.1 Experimental Design and Procedure. The hypothesis behind this experiment posited that
viewing differences would exist between those who recognized the painting and those who did not.
The epxeriment proceeded with participants viewing the stylized (abstracted) representation for 7
Artwork attractiveness
high
low
0.6
Transition entropy
0.4
0.2
0.0
neutral positive
Experimental condition
Fig. 6. Transition entropy depending on subjective attractiveness of artwork and experimental condition. Error bars represent
1SE.
(a) La Jaconde (b) La Jaconde stylized
Fig. 7. Painting and its point-like representation.
Fixations Fixations
896 896
768 768
y-coordinate (pixels)
y-coordinate (pixels)
640 640
512 512
384 384
256 256
128 128
0 0
0 160 320 480 640 800 960 1120 0 160 320 480 640 800 960 1120
x-coordinate (pixels) x-coordinate (pixels)
(a) Viewer who recognized La Jaconde (b) Viewer who did not recognize La Jaconde
seconds (called puntos1 image). They pressed the space bar as soon as they recognized the painting.
They then viewed the digital representation of the painting in its original form (called the picture
image), followed once again by the abstracted version (called puntos2 this time). Differences in visual
behavior between those who recognized the abstract version of the painting would thus be expected in
gaze data collected over puntos1.
6.3.2 Participants. Thirty-one (31) participants took part in the study. Of those, 17 self-reported
that they did not recognize the abstract version of the painting, while 14 claimed they did.
6.3.3 Apparatus. Gaze data was recorded by a Tobii T60 (60 Hz) eye tracker while participants
viewed a 17 inch monitor (set to 1280 1024 resolution) at an approximate viewing distance of 19.68
inches.
Raw data collected from the tracker was smoothed with a Butterworth filter set to 60 Hz and 5.65
Hz sampling and cutoff frequencies, respectively. Smoothed data was then convolved with a Savitzky-
Golay differential filter acting as a velcity-based saccade detector (7-tap filter width with velocity
threshold set to 5 /s.
6.4 Results
Analysis of the eye movement data focuses on comparison of subjects transition and stationary en-
tropies, in relation to their self-reported identification of the abstracted stimlus (puntos1) and the
granularity of the AOI grid.
6.4.1 Transition and Stationary Entropies. When using an 8 8 AOI grid, a one-way repeated-
measures ANOVA on normalized transition entropy between the three image viewing tasks fails to
show a statistical difference in gaze transition (F (2, 58) = 0.81, p = 0.45, n.s., see Fig. 8(c)). Mean nor-
malized entropy of transition entropy over the picture (Ht = 0.71, SD = 0.15) is greater than the
normalized entropy for puntos1 (Ht = 0.70, SD = 0.13) and the normalized entropy for puntos2
(Ht = 0.67, SD = 0.13), but not significantly so.
A one-way repeated-measures ANOVA fails to reveal statistical signfiance between mean stationary
entropies, (F (2, 58) = 0.78, p = 0.46, n.s., see Fig. 8(c)). Although mean normalized stationary entropy
of transitions over the picture is greater or equal (Hs = 0.91, SD = 0.06) to the normalized stationry
Normalized Entropy vs. Task Type Normalized Entropy vs. Task Type
stationary entropy stationary entropy


transition entropy
1.0
1.0
transition entropy
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
picture puntos1 puntos2 no si
Stimulus Painting Recognition
(c) Entropies vs. task (d) Entropies vs. recognition
Fig. 8. Transition and stationary entropies vs. stimulus viewing task or vs. recognition.
entropses for either puntos1 (Hs = 0.91, SD = 0.05) or puntos2 (Hs = 0.89, SD = 0.06), this difference is
not significant.
Statistical comparison of entropy of gaze transitions shows no significant differences in viewing
behavior when looking at the two stimulus images, when modeled as first-order Markov stochastic
processes. Due to higher entropy, results sgugest that transitions are less predictable over the familiar
image than over the abstracted image of La Jaconde.
Using the same 88 AOI grid, a Welch two sample t-test on normalized transition entropy between
the two viewer groups, those who either self-reported recognizing the abstracted painting or those
who did not, shows a statistical difference in gaze transition (t = 3.67, p < 0.01, see Fig. 8(d)). Mean
normalized entropy of viewers who did not recognize La Jaconde (Ht = 0.64, SD = 0.12) is lower than the
normalized entropy for those who did (Ht = 0.78, SD = 0.09).
A Welch two sample t-test reveals statistical signfiance between mean stationary entropies, as well,
(t = 3.96, p <!0.01, see Fig. 8(d)). Mean normalized stationary entropy of viewers who did not recog-
nize La Jaconde (Hs = 0.88, SD = 0.05) is lower than the normalized stationry entropy for those who
did (Hs = 0.94, SD = 0.03), and this difference is statistically significant. This suggests that viewers
less predictable gaze transitions atop the familiar image of La Jaconde carries on into the long term
(steady-state).
6.4.2 Effect of Grid Granularity. When using a finer 108 AOI grid, a one-way repeated-measures
ANOVA on normalized transition entropy between the three image viewing tasks fails to show a statis-
tical difference in gaze transition (F (2, 58) = 2.98, p < 0.06, n.s.). Mean normalized entropy of transition
entropy over the picture (Ht = 0.23, SD = 0.06) is similar to the normalized entropy for puntos1
(Ht = 0.24, SD = 0.05) and the normalized entropy for puntos2 (Ht = 0.26, SD = 0.04).
Statistical comparison of entropy of gaze transitions reveals no significant differences in viewing

behavior when looking at the two stimulus images.
A finer 108 AOI grid also leads to no signficant differences between normalized transition entropies
of the two viewer groups, those who either self-reported recognizing the abstracted painting or those
who did not, according to a Welch two sample t-test (t = 2.00, p = 0.05, n.s.). Mean normalized entropy of
viewers who did not recognize La Jaconde (Ht = 0.26, SD = 0.04) is statistically similar to the normalized
entropy for those who did (Ht = 0.22, SD = 0.06).
Reasons for lack of significance are again likely due to the coarseness of the AOI grid used in this
analysis.
7. DISCUSSION AND CONCLUSIONS

We introduced a new approach to the analysis of eye movement transitions between different Areas
of Interest. The approach is based on Markov chain modeling of individual switching patterns. Besag
and Mondals [2013] statistical test confirms the validity of modeling transitions as short, first-order
(one fixation memory) processes. The complexity of transition distributions between AOIs and the com-
plexity of the resultant stationary distribution can be expressed in terms of Shannons entropy. Higher
transition entropy Ht denotes more randomness and more frequent switching between AOIs. As a re-
sult, individual switching patterns between AOIs, which can vary significantly between subjects, can
be visualized and statistically quantified. On the other hand, greater stationary entropy Hs means that
visual attention is distributed more equally among AOIs, perhaps because they hold similar interest
for the viewer. Stationary entropy, therefore, can be thought of as a measure of interest, although one
should consider adjusting stationary probabilities due to different AOI sizes.
The two-step approach to analysis of eye movement transitions was successfully demonstrated on
data from a group of subjects who were asked to freely explore three examples of classical artwork.
Results showed that transition entropy can capture individual differences. Subjects deemed more cu-
rious demonstrated significantly more local and predictable patterns of visual scanning, resulting in
lower transition entropies compared to those deemed less curious. Moreover, transition and stationary
entropies were related to subjective aesthetic evaluation of the art. Participants who better appreci-
ated the artwork were characterized by lower transition entropy and higher stationary entropy. We
attribute these findings to more deliberate and focused visual investigation.
Further development of the analytic method should consider its limitation of pre-defined number
and size of AOIs for a given stimulus. The value of entropy depends on the number of AOIs. The more
AOIs, the greater the entropy. Bearing this in mind, standardization of both transition and station-
ary entropy coefficients is required. Too many AOIs may result in a significant number of cells in the
transition matrix with empty cells. By increasing the number of AOIs, the transition and stationary
probabilities (and therefore the entropies) will be estimated with a larger error. Furthermore, station-
ary distribution (and so its entropy) may not exist or may not be unique. The entropy can be calculated
from such data, but the results might be misleading. Therefore, for now we suggest to choose 2-4 AOIs
(following Ponsoda et al. [1995]) and to ensure that stimuli presentation time is sufficient for dense
transition matrices.
Future work should address three issues related to the number and size of AOIs. First, the influ-
ence of sparse matrices on interpretation of entropy coefficients. Second, the dependence of stationary
entropy on the size of AOIs. Third, standardization and determination of the optimal number of AOIs
for different types of stimuli. In principle, using fixated AOI scanpath indices, i.e., the content-driven
analysis approach, the method could also be extended to dynamic stimuli provided that the set of AOIs
S = {1, . . . , s} (the state space) is fixed. The state space can operationally be fixed either over the entire
exposure duration of the stimulus (if this makes sense, e.g., scene vs. subtitle transitions when view-
ing subtitled film), or over a number of small temporal segments of the stimulus, e.g., constituting a
frame-based approach as suggested by Grindinger et al. [2010a]. However, if the underlying stimulus
changes dynamically, then the likelihood of fixating any given AOI is likely to change, which would
affect the expected viewing probabilities over time. Thus, a stimulus-driven analysis would need to
consider the dynamically changing fixation probabilities as they are likely to change in response to
the stimulus. For example, one could segment temporally more or less based on the constancy of the
underlying stimulus, but this would be limited by the amount of fixations available per segment.
To conclude, the proposed method is fairly straightforward to implement with any scripting language
and does not require extensive computational resources. The most time consuming part of the method
is the Markov chain goodness-of-fit statistical test. The testing procedure (written in R), for every
stimulus and subject, takes about 1 minute to produce the p-value (running on an Intel multi-core PC-
class computer). Modeling eye movement transitions between AOIs as Markov chains and calculating
transition and stationary entropies produces straightforward measures of the predictability of saccadic
transitions atop stimuli. These measures can be tested for individual differences due to individual
characteristics or stimuli properties via popular statistical tests.
REFERENCES
A CART URK , C. AND H ABEL , C. 2012. Eye tracking in multimodal comprehension of graphs. In TEDR-2012: Proceedings of the
Workshop on Technology-Enhanced Diagrams Research, R. Cox, Ed. 1125.
A LTHOFF , R. R. 1998. Eye movement-based memory assessment: The use of eye movement monitoring as an indirect measure
of memory. Ph.D. thesis, University of Illinois at Urbana-Champaign.
B EDNARIK , R., M YLLER , N., S UTINEN, E., AND T UKIAINEN, M. 2005. Applying eye-movement tracking to program visualiza-
tion. In VLHCC 05: Proceedings of the 2005 IEEE Symposium on Visual Languages and Human-Centric Computing. IEEE
Computer Society, Washington, DC, USA, 302304.
B ESAG, J. AND M ONDAL , D. 2013. Exact goodness-of-fit tests for Markov chains. Biometrics 69, 2, 488496.
B IELE , C., K OPACZ , A., AND K REJTZ , K. 2013. Shall we care about the users feelings? Influence of affect and engagement on
visual attention. In Proceedings of the International Conference on Multimedia, Interaction, Design and Innovation. MIDI 13.
ACM, New York, NY, USA, 7:17:8.
C IUPERCA , G. AND G IRARDIN, V. 2005. On the estimation of the entropy rate of finite Markov chains. In Proceedings of the
XIth International Symposium on Applied Stochastic Models and Data Analysis. Brest, France.
DE B RUIN, J. A., M ALAN, K. M., AND E LOFF , J. H. P. 2013. Saccade deviation indicators for automated eye tracking analysis.
In Proceedings of the 2013 Conference on Eye Tracking South Africa. ETSA 13. ACM, New York, NY, 4754.
D UCHOWSKI , A. T., D RIVER , J., J OLAOSO, S., T AN, W., R AMEY, B. N., AND R OBBINS, A. 2010. Scanpath Comparison Revisited.
In Proceedings of the 2010 Symposium on Eye-Tracking Research & Applications. ETRA 10. ACM, New York, NY, 219226.
E KROOT, L. AND C OVER , T. M. 1993. The Entropy of Markov Trajectories. IEEE Transactions on Information Theory 39, 4,
14181421.
F ISCHER , P. AND P EINSIPP -B YMA , E. 2007. Eye Tracking for Objective Usability Evaluation. In European Conference on Eye
Movements (ECEM). ECEM.
G OLDBERG, J. H. AND K OTVAL , X. P. 1999. Computer Interface Evaluation Using Eye Movements: Methods and Constructs.
International Journal of Industrial Ergonomics 24, 631645.
G RINDINGER , T., D UCHOWSKI , A. T., AND S AWYER , M. 2010a. Group-Wise Similarity and Classification of Aggregate Scan-
paths. In ETRA 10: Proceedings of the 2010 Symposium on Eye Tracking Research & Applications. ACM, New York, NY,
101104.
G RINDINGER , T. J., M URALI , V. N., T ETREAULT, S., D UCHOWSKI , A. T., B IRCHFIELD, S. T., AND O RERO, P. 2010b. Algorithm
for Discriminating Aggregate Gaze Points: Comparison with Salient Regions-Of-Interest. In International Workshop on Gaze
Sensing and Interactions (IWGSI; published in ACCV 2010 Workshops, Part I, LNCS 6468). IWGSI/ACCV, Queenstown, New
Zealand, 390399.
H AUSSER , J. AND S TRIMMER , K. 2009. Entropy Inference and the James-Stein Estimator, with Application to Nonlinear Gene
Association Networks. Journal of Machine Learning Research 10, 14691484.
H UME , E. AND M AILHOT, F. 2013. The Role of Entropy and Surprisal in Phonologization and Language Change. In Origins of
Sound Change: Approaches to Phonologization, A. C. L. Yu, Ed. Oxford University Press, Oxford, UK, 2947.
H WANG, A. D., WANG, H.-C., AND P OMPLUN, M. 2011. Semantic guidance of eye movements in real-world scenes. Vision
Research 51, 11921205.
J ARODZKA , H., H OLMQVIST, K., AND N YSTR OM , M. 2010. A vector-based, multidimensional scanpath similarity measure.
In Proceedings of the 2010 Symposium on Eye-Tracking Research and Applications. ETRA 10. ACM, New York, NY, USA,
211218.
K ASHDAN, T. B., G ALLAGHER , M. W., S ILVIA , P. J., W INTERSTEIN, B. P., B REEN, W. E., T ERHAR , D., AND S TEGER , M. F.
2009. The curiosity and exploration inventory-ii: Development, factor structure, and psychometrics. Journal of Research in
Personality 43, 987998.
K REJTZ , K., S ZMIDT, T., D UCHOWSKI , A. T., AND K REJTZ , I. 2014. Entropy-based Statistical Analysis of Eye Movement
Transitions. In Proceedings of the 2014 Symposium on Eye-Tracking Research and Applications. ETRA 14. ACM, New York,
NY.
K RUIZINGA , A., M ULDER , B., AND DE WAARD, D. 2006. Eye scan patterns in a simulated ambulance dispatchers task. In
Developments in Human Factors in Transportation, Design, and Evaluation. Shaker Publishing, Orbassano, Italy.
K ULLBACK , S. AND L EIBLER , R. 1951. On information and sufficiency. Annals of Mathematical Statistics 22, 7986.
L IECHTY, J., P IETERS, R., AND W EDEL , M. 2003. Global and Local Covert Visual Attention: Evidence From a Bayesian Hidden
Markov Model. Psychometrika 68, 4, 519541.
P ERILLI , F. G., P IRES, A., V ILAR O , A., V ILLALOBOS, N., C APURSO, F., G O NI , R., F ERNANDEZ , E., AND D UCHOWSKI , A.
2014. Recognizing Mona Lisa: An Eye-Tracker Study. In XVII Congreso de la Sociedad Iberoamericana de Grafica Digital
(SIGRADI). (In review).
P ONSODA , V., S COTT, D., AND F INDLAY, J. M. 1995. A probability vector and transition matrix analysis of eye movements
during visual search. Acta Psychologica 88, 167185.
R D EVELOPMENT C ORE T EAM. 2011. R: A Language and Environment for Statistical Computing. R Foundation for Statistical
Computing, Vienna, Austria. ISBN 3-900051-07-0.
R ICHELL , R. AND A NDERSON, M. 2004. Reproducibility of negative mood induction: A self-referent plus musical mood induction
procedure and a controllable/uncontrollable stress paradigm. J. Psychopharmacol. (Oxford) 18, 1, 94101.
R OBINSON, O., G RILLON, C., AND S AHAKIAN, B. 2012. The mood induction task: A standardized, computerized laboratory
procedure for altering mood state in humans. Protocol Exchange.
S HIC, F., C HAWARSKA , K., B RADSHAW, J., AND S CASSELLATI , B. 2008. Autism, eye-tracking, entropy. In Development and
Learning, 2008. ICDL 2008. 7th IEEE International Conference on. 7378.
VANDEBERG, L., B OUWMEESTER , S., B OCANEGRA , B. R., AND Z WAAN, R. A. 2013. Detecting cognitive interactions through
eye movement transitions. Journal of Memory and Language 69, 445460.
V ELICHKOVSKY, B. M., J OOS, M., H ELMERT, J. R., AND PANNASCH , S. 2005. Two Visual Systems and Their Eye Movements:
Evidence from Static and Dynamic Scene Perception. In CogSci 2005: Proceedings of the XXVII Conference of the Cognitive
Science Society. Stresa, Italy, 22832288.
V ELTEN, E. 1968. A laborator task for induction of mood states. Behav Res Ther 6, 4, 473482.
W OBBROCK , J., F INDLATER , L., G ERGLE , D., AND H IGGINS, J. 2011. Aligned rank transform for nonparametric factorial
analyses using only anova procedures. In CHI 2011. 143146.

Gaze Transition Entropy, KRZYSZTOF KREJTZ, ANDREW DUCHOWSKI

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gaze Transition Entropy, KRZYSZTOF KREJTZ, ANDREW DUCHOWSKI

Uploaded by

Copyright:

Available Formats

1

Gaze Transition Entropy

3. MARKOV PROPERTY OF GAZE TRANSITIONS

3.1 Testing of the Markov Property

4. ENTROPY OF AOI SEQUENCE

of the Jensens inequality1

5.3 Stationary Entropy Implementation via Eigen Analysis

6. EXAMPLE APPLICATIONS OF TRANSITION ENTROPY ANALYSIS

(a) Renaissance2 AOIs (b) Renaissance3 AOIs

2 The picture of the artwork was retrieved from http://en.wikipedia.org/wiki/File:Giorgione_019.jpg. Source/photogra-

(a) Ht = 0.89, Hs = 1.00 (b) Ht = 0.69, Hs = 0.98 (c) Ht = 0.71, Hs = 0.78

1 0.52 0.39 0.10 1 0.71 0.21 0.08 1 0.33 0.53 0.13

Source AOI (from)

Source AOI (from)

3 0.15 0.21 0.64 3 0.03 0.19 0.77 3 0.05 0.47 0.47

(d) Ht = 0.89 (e) Ht = 0.69 (f) Ht = 0.71

Table I. P-values of the exact (asymptotic) Markov chain

3 We used testing procedure in R written by Besag and Mondal [2013].

Renaissance 2 AOI Renaissance 3 AOI

Renaissance 2 AOI Renaissance 3 AOI

Mean Normalized Entropy (bits/transition with SE)

Entropy type Entropy type

Bauhaus Impressionism Renaissance Bauhaus Impressionism Renaissance

Bauhaus Impressionism Renaissance

6.3 Viewing La Jaconde

(a) La Jaconde (b) La Jaconde stylized

Fig. 7. Painting and its point-like representation.

stationary entropy stationary entropy

Mean Normalized Entropy (bits/transition with SE)

(c) Entropies vs. task (d) Entropies vs. recognition

Statistical comparison of entropy of gaze transitions reveals no significant differences in viewing

7. DISCUSSION AND CONCLUSIONS

You might also like