Master Thesis - Algorithm for pattern recognition

Hochschule Fulda
Fachbereich Angewandte Informatik

University of Massachusetts Boston
Department of Computer Science
Visual Attention Laboratory

Master of Science Thesis

Master of Science of Electronic Business

An algorithm for pattern recognition driven by eye
movement

by

Andreas Lennartz

Supervisors:
Prof. Dr. Jan-Torsten Milde
Prof. Dr. Marc Pomplun

Boston, September 2007

Abstract

A lot of different works were published which examine basic search task where
a user has to find a given object inside some picture or other object. During this
task the subject’s eye movement are recorded and later on analyzed for a better
understanding of a human’s brain and the corresponding eye movements to a
given task. In such search tasks like ”find-the-object” the question arises if it is
possible to determine what a subject is looking for just by considering the given
eye movement data, without knowing what he/she is looking for.
In this work an eye tracking experiment was introduced and conducted. The
experiment presented different random-dot pictures to the subjects, consisting
of squares in different colors. In these pictures the task was to find a pattern
with a size of 3x3 squares. For the first part of the experiment, the used squares
were in black and white, in the second part gray was added as an additional
color. During each experiment the eye movements were recorded.
Special software was developed and introduced to convert and analyze the
recorded eye movement data, to apply an algorithm and generate reports that
summarize the results of the analyzed data and the applied algorithm.
A discussion of these reports shows that the developed algorithm works well
for 2 colors and different square sizes used for search pictures and target pat-
terns. For 3 colors it is shown that the patterns which the subjects are searching
for are too complex for a holistic search in the pictures - the algorithm gives poor
results. Evidences are given to explain this results.

Contents

Contents i

1 Introduction 3
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Chapter Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Visual Search and Eye Tracking 5
2.1 Human Vision and Visual Search . . . . . . . . . . . . . . . . . . . 5
2.1.1 Human Vision . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Visual Search . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Tracking Eye Movements . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Deﬁnition Eye Tracking . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Eye Link II from SR Research . . . . . . . . . . . . . . . . . 7
2.3 Visual Search Experiments . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Experiment Description . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Experiment Software 17
3.1 Experiment Builder . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.1 Step by Step Experiment Builder . . . . . . . . . . . . . . . 17
3.1.2 Experiment Builder Files . . . . . . . . . . . . . . . . . . . . 23
3.1.3 Class Overview . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Experiment Engine . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1 Step by Step Experiment Engine . . . . . . . . . . . . . . . 25
3.2.2 Code Overview . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Experiment Analysis Tool . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.1 Step by Step Experiment Analysis Tool . . . . . . . . . . . 29
3.3.2 Analysis Output Files . . . . . . . . . . . . . . . . . . . . . 33
3.3.3 Class Overview . . . . . . . . . . . . . . . . . . . . . . . . . 42

4 Experiment Setup, Run and Analysis 47
4.1 Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

i

ii CONTENTS

4.1.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.1.2 Apparatus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.1.3 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Experiment Runs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3 Experiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.1 Reports Experiment I . . . . . . . . . . . . . . . . . . . . . . 50
4.3.2 Discussion Experiment I . . . . . . . . . . . . . . . . . . . . 64
4.3.3 Reports Experiment II . . . . . . . . . . . . . . . . . . . . . 65
4.3.4 Discussion Experiment II . . . . . . . . . . . . . . . . . . . 73

5 Conclusion 75

Bibliography 79

A Used software and CD contents 83
A.1 Used Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
A.2 Contents of the Enclosed CD . . . . . . . . . . . . . . . . . . . . . 84

B Namespace ASCLibrary 85
B.1 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
B.1.1 Class ASCBlink . . . . . . . . . . . . . . . . . . . . . . . . . 86
B.1.2 Class ASCData . . . . . . . . . . . . . . . . . . . . . . . . . 87
B.1.3 Class ASCFixation . . . . . . . . . . . . . . . . . . . . . . . 88
B.1.4 Class ASCSaccation . . . . . . . . . . . . . . . . . . . . . . 90
B.1.5 Class ASCTrialData . . . . . . . . . . . . . . . . . . . . . . 92
B.1.6 Class ASCTrialMetaData . . . . . . . . . . . . . . . . . . . 94

C Namespace BMPAnalysis 97
C.1 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
C.1.1 Class AnalysisObject . . . . . . . . . . . . . . . . . . . . . . 99
C.1.2 Class ColorDistribution . . . . . . . . . . . . . . . . . . . . 101
C.1.3 Class Conﬁg . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
C.1.4 Class Info . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
C.1.5 Class MainForm . . . . . . . . . . . . . . . . . . . . . . . . 106
C.1.6 Class OutputFile . . . . . . . . . . . . . . . . . . . . . . . . 107
C.1.7 Class OutputFileExcel . . . . . . . . . . . . . . . . . . . . . 110
C.1.8 Class OutputFileParameter . . . . . . . . . . . . . . . . . . 111
C.1.9 Class OutputFileRawData . . . . . . . . . . . . . . . . . . . 114
C.1.10 Class Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

D Namespace BMPBuilder 123
D.1 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
D.1.1 Class Bitmap . . . . . . . . . . . . . . . . . . . . . . . . . . 125

iii

D.1.2 Class Constants . . . . . . . . . . . . . . . . . . . . . . . . . 128
D.1.3 Class ExperimentData . . . . . . . . . . . . . . . . . . . . . 133
D.1.4 Class ExperimentForm . . . . . . . . . . . . . . . . . . . . 135
D.1.5 Class Info . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
D.1.6 Class Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . 137
D.1.7 Class Picture . . . . . . . . . . . . . . . . . . . . . . . . . . 138
D.1.8 Class PreviewData . . . . . . . . . . . . . . . . . . . . . . . 141
D.1.9 Class PreviewForm . . . . . . . . . . . . . . . . . . . . . . 145
D.1.10 Class TrialData . . . . . . . . . . . . . . . . . . . . . . . . . 146
D.1.11 Class TrialForm . . . . . . . . . . . . . . . . . . . . . . . . . 147

List of Symbols and Abbreviations 149

List of Figures 150

List of Tables 152

Acknowledgements

Thanks to my parents for the mental and ﬁnancial support.
Thanks to my supervising professors Dr. Marc Pomplun and Dr. Jan-Torsten
Milde for their scientiﬁc support.

1

Chapter 1

Introduction

1.1 Motivation

A lot of research is done about the human recognition of patterns or snippets
in pictures. For this purpose the movement of the gaze is tracked with special
hardware (namely eye tracker), while the subject tries to find a snippet of a given
picture. The subject searches for this snippet in the picture until he/she recog-
nizes it, meanwhile eye movements are recorded. Analyzing the eye movement
of these records afterward, it is obvious that only special parts of the picture are
interesting for the subject, other parts are not. There is research done that tries
to determine reasons for this, e.g. correlations between the color of the snippet
and similar colors at the observed parts. The question arises if it is possible to
determine the snippet that the subject was searching for, without knowing the
snippet itself but the parts the subject has looked at.
The goal of this thesis is to determine a possible algorithm that takes eye
movement input data of observed pictures in which the subject were to find a
pattern. The eye movement data is analyzed and one or more possible patterns
that could be the search pattern are given. In order to do this an experiment will
be constructed where different random-dot pictures are given and the eye move-
ments of the subjects looking for the search pattern in the picture are recorded.
Then one algorithm will be developed, where existing research is considered.
This algorithm will be tested for its functionality. Because of the great amount
of possibilities the random-dot pictures (and their patterns) should be of a kind
which can be easily processed by computers (e.g. black and white squares). Such
an algorithm for detailed estimation of a user’s intent would be valuable for the
development of more efficient human-computer interfaces.

3

4 CHAPTER 1. INTRODUCTION

1.2 Chapter Overview
Chapter 2 introduces different aspects of visual search and eye tracking. Section
2.1 gives a brief explanation of the human vision and how human beings con-
duct visual search, as well as the importance of visual search in scientiﬁc studies.
Section 2.2 explains how eye movements can be tracked and recorded using an
eye tracking device. Existing research in the ﬁeld of visual search is discussed
in section 2.3. In order to create an algorithm that determines a target pattern
out of eye movements, and based on the problems with current research, a new
experiment is created for this work and described in section 2.4.
Chapter 3 explains the software developed and how it is used to build, run
and analyze the newly designed eye tracking experiment. For this purpose three
software parts were developed: An Experiment Builder to build experiments,
described in section 3.1, an Experiment Engine to run an experiment, described
in section 3.2, and an Experiment Analysis Tool to analyze the test data and
create human-readable reports, described in section 3.3.
Chapter 4 explains the results of the experiment itself and how the devel-
oped algorithm works in different tests. It contains the experiment setup in sec-
tion 4.1, the experiment runs in section 4.2 and the experiment results in section
4.3.
Chapter 5 is the conclusion that summarizes and discusses the results of this
work.

Chapter 2

Visual Search and Eye Tracking

2.1 Human Vision and Visual Search

2.1.1 Human Vision

The human eye is a very complex system. There is a region called fovea in the
center of the retina which has a high density of photo receptors and therefore
a high resolution. Outside this area (with a radius of about one degree of vi-
sual angle) the density of photo receptors decreases quickly, so the resolution
is lower. This means that we have a very detailed vision in the center of our
visual field and only coarse perception in the outer regions of the eye. For this
reason the human eye is in a constant movement to receive a good vision of the
environment.
Eye movements are not performed in a smooth manner like it perhaps can
be assumed. The gaze position ”jumps” between inspected locations. These
quick jumps are called saccades and the motionless phases in between are called
fixations. Visual information can only be perceived during fixations, according
to the brain’s ability of constructing a stable image of the surroundings.[24]
A fixation duration can be considered as a measure of the effort to process
information. The longer our attention rests at a certain location, the longer it
presumably takes to deal with the presented visual information. (Eye-mind hy-
pothesis, [7])
The saccade length reveals how thoroughly a certain region of the visual field
is scanned. Short saccades indicate that the fovea is directed to positions close
to each other, which means a high resolution scanning process. Long saccades
imply that the local scene is only roughly perceived or that its information con-
tent is low. Fixation duration and saccade length are the basic eye-movement
variables. Additionally, eye movements yield data about where and in which

5

6 CHAPTER 2. VISUAL SEARCH AND EYE TRACKING

temporal order a subject acquires visual information, e.g. eye movements reveal
a subject’s distribution and dynamics of visual attention.[12, p.5]

2.1.2 Visual Search
The visual search is a common and in the everyday live very important task
for the human being. The success of localization of objects in a given area, e.g.
the never-ending search for your keys, is determined by the active nature of
looking. Because the human vision is a ”foveated” visual system, it must be
able to combine several eye movements to scan a visual field.
Most information during a visual search task are gathered during the fixa-
tions, and only little during the saccades, due to saccadic suppression and mo-
tion blurring. During each fixations in a search task, the scene is scanned for
the target with the center of the eye for a high-resolution image, and the low-
resolution periphery is used to plan the next subsequent fixation. An under-
standing of how the human visual system selects images could be important for
a better understanding of the biological vision as well as a fundament for any
artificial vision or human-computer interaction. ([2])

2.2 Tracking Eye Movements
In order to examine the eye movements of a human being the field of tracking
eye movements (namely eye tracking) was introduced.

2.2.1 Definition Eye Tracking
As a very common entry point to gain quick knowledge of specific topics, the on-
line editable-by-all dictionary Wikipedia is always a good starting point (though
the information should always be double checked for its validity):
”Eye tracking is the process of measuring either the point of gaze (where we
are looking) or the motion of an eye relative to the head. There are a number of
methods for measuring eye movements. An instrument that does eye tracking
is called an eye tracker.” [21]
Eye tracking is part of the cognitive sciences, which resulted from the rea-
sonable view that in order to understand the human brain scientists of different
faculties like psychology, medicine and computer science should work together
in interdisciplinary teams. These teams try to combine their knowledge and
abilities to explore the brain’s mechanism. [12, p.1]
In order to track and record the eye movements, special hardware is needed.
These hardware is commonly called EyeTracker, and such a device used for this
work is introduced in the next section.

2.2. TRACKING EYE MOVEMENTS 7

Figure 2.1: EyeLink II eye tracking device

2.2.2 Eye Link II from SR Research

The Laboratory at University of Massachusetts Bostons uses such an Eye Track-
ing Device, namely EyeLink II from SR Research.
The EyeLink II system consists of three miniature cameras mounted on a
padded headband. Two eye cameras allow binocular eye tracking or monocular
by selecting of the subject’s dominant eye. These cameras have built-in illumi-
nators to even lighting for the entire ﬁeld of view. Also combined with digital
compensation for changes in ambient lighting, this results in stable pupil acqui-
sition. (Figure 2.1)
The EyeLink Tracker from SR Research has a high resolution with a low noise
level and a fast data rate (up to 500 samples per second, which means 500Hz)
with the mounted video-based eye tracker. This data quality results in very low
velocity noise, making the EyeLink II useful for saccade analysis and smooth
pursuit studies. In addition, on-line gaze position data is available with delays
as low as 3 milliseconds, making the system ideal for gaze-contingent display
applications. [16]
The third camera, an optical head-tracking camera integrated into the head-
band, allows accurate tracking of the subject’s point of gaze. The low noise level
of this optical tracker means that the noise level of computed gaze position data
is comparable to that of the original eye position data, allowing gaze data to be
used for saccadic analysis. In contrast, the magnetic head trackers used in other
head-mounted eye tracking systems have high angular noise and therefore limit
the usability of gaze position data.
EyeLink II has a high compatibility with a wide range of subjects. Dark


Figure 2.2: EyeLink II connection schema

pupil tracking and off-axis illumination allow tracking of subjects with almost
all eyeglasses, eliminating the bright reflections. Eye camera sensitivity is high
enough that even eyeglasses with heavy anti-reflection coatings that block up to
80% of infrared illumination are not a problem.
The EyeLink II eye tracking device (in the following called EyeTracker) is
connected with its EyeLink II Software running on its host PC. The host PC can
transfer data to the display PC, usually the computer where there experimental
stimuli is displayed. Also a control device such as a game pad can be connected
to the host (and/or display) PC. Both display and host PC are connected with a
standard Ethernet cable. The monitor screen of the display PC has a set of mark-
ers attached to it, which are delivering additional information of the subjects’
head position to the EyeTracker. (See figure 2.2).
Figure 2.3 shows an example for the camera setup screen of the EyeLink II
software.

2.3 Visual Search Experiments
In order to examine the visual search by humans, experiments using eye track-
ing systems such as the EyeTracker were created. In these experiments which ex-
amined the detection of targets and involved distracters, first of all it was found
out that the saccadic movement during a visual search task can be influenced
and guided by visual cues ([23]). These cues can be of different types. The most
important characteristics of these cues are color ([22]) and shape ([5]). These
visual cues can affect the search performance ([20]). However, the experiments
do not reveal what image feature or structural cues are attracting the observer’s

2.3. VISUAL SEARCH EXPERIMENTS 9

Figure 2.3: EyeLink II camera setup screen shot

gaze.
Eye movements are crucial for any (foveated) visual system as a salient fea-
ture of visual search. There has been considerable investigations in free viewing
studies to examine what kind of image features influence human eye move-
ments. Free viewing studies analyze the eye movements of a human subject
without a given task. An analysis at the point of gaze shows that eye move-
ments are not really random - they are possibly drawn to image patches with a
high contrast and low interpixel correlation. ([11],[15])
The question is how such free viewing experiments can be correlated to vi-
sual search problems - if a search task is given in first place, the selectivity of the
eye movements is strongly influenced. Therefore, free viewing studies are not
ideal to examine visual search. In studies with a given task, it could be examined
that eye movements and especially the fixations of the eye try to analyze parts
of the stimuli which are similar to the target. This means that we peripherally
select areas which are similar looking (or are the same) to the search target, and
then our fixation will analyze if the selected area really correlates to the target.
[13]
Based on these perceptions, research has been done to prove that eye move-
ment is depending on the search task and the stimuli. In an experiment by Ra-
jashekar, Bovik & Cormack (see [14]) the eye movement of three observers were
recorded while they tried to find a simple geometric shape (circle, dipole, tri-
angle, or bow tie) in a ”noise stimuli” picture. The shapes are shown in figure
2.4. The stimuli were based on a Fourir spectrum where the noise matched the
average spectrum of natural images. An example stimuli with observer’s eye
movements superimposed is seen in figure 2.5. A quick examination of the stim-
uli might show the difficulty of the search task. Further it could be assumed
that in such a low signal-to-noise ratio stimuli observers might be simply de-
ploy their eye movement uniformly across the stimuli. But the analysis showed


Figure 2.4: Simple geometric target shapes, [14]

Figure 2.5: Noise stimuli, with superimposed eye movements from a trial, [14]

that a different strategy is applied.
For the analyzation of the experiment, a classification image technique was
used by incorporating eye movements as described in the following. For each
subject and target, a region with the size 128x128 pixel (twice the size of a target)
around each fixation was used. These regions were averaged across the trials
to yield classification images. All fixations by the subjects, including the final
fixation and those that landed on the target, were used in the analysis. In order
to enhance the details in the classification images, a statistical thresholding was
performed by setting pixel intensities within one standard deviation of the mean
equal to the mean (referring to [1]).
The statistically thresholded results for all subjects (namely LKC, EM and
UR) and targets (circle, dipole, triangle, and bow tie) are shown in figure 2.6.
The first row illustrates the four targets that observers were instructed to find in
the stimuli. Each of the other rows shows the classification images for the three
observers for these targets.
In order to quantify the uncertainty associated with the shapes of these clas-
sification images, the authors bootstrapped∗ the averaging procedure and com-

∗
Bootstrapping is the practice of estimating properties of an estimator (such as its variance) by
measuring those properties when sampling from an approximating distribution. One standard

2.3. VISUAL SEARCH EXPERIMENTS 11

Figure 2.6: Statistically thresholded results for all subjects, [14]

puted the boundaries of each of the resulting thresholded classification images
as follows. To detect the boundaries of the statistically filtered classification im-
ages, the two largest regions in the statistically filtered image were detected us-
ing a connected components algorithm. The outlines of each of these regions
were then used to represent the boundary of the classification image. Bootstrap-
ping was then used to verify the robustness of these boundaries. The boundaries
of the resulting classification images were then added together and superim-
posed on the classification images to reflect the stability of the boundary. The
aggregate of all of the bootstrapped boundaries is shown superimposed in red
in figure 2.6.
The red well-defined boundary indicates that the structure in the classifica-
tion image is robust, and diffused edges indicate that the shape of the classifi-
cation image in that region varies across trials. Rajashekar, Bovik, & Cormack
argue that the results show that subjects fixated areas which look like the target
shape. The results should prove that there is a computable connection between
the fixations and the target object, and that subject’s fixations are tending to look
at areas in the stimuli which are highly similar to the target shape. But figure 2.6
on which these assumptions are based shows that a qualitatively proof is miss-
ing.
Summarizing, research shows that there is a connection between the given

choice for an approximating distribution is the empirical distribution of the observed data. [4]


search task and the subject’s eye movement. Regarding Rajashekar, Bovik, &
Cormack ([14]), and also Pomplun ([13]), it is likely that a subject’s gaze tends
to ﬁxate shapes in a stimuli that are similar to the target shape. But how much
information does the correlation between eye movement and target shape really
contain? Is it possible to ﬁnd a measurable way to qualitatively examine this
information? The question arises if it possible to determine the target shape by
analyzing only the subject’s eye movement, where a priori the target shape is
unknown. This would give valuable statements about the connection between
the target and the stimuli. Based on this, the following experiment for an Eye-
Tracker device was created.

2.4. EXPERIMENT DESCRIPTION 13

Figure 2.7: Example of target pattern and search picture with 2 colors

2.4 Experiment Description
For this work an eye tracker experiment is supposed to run with several subjects.
The subjects’ task is to search a random-dot pattern consisting of 3x3 squares in
a stimuli picture. The number of colors are either 2 or 3 (white, black and gray as
additional color). Figure 2.7 shows an example of the target pattern and stimuli
picture for 2 colors, figure 2.8 for three colors. The search task will be repeated
30 times for each subject, each time with the same target pattern but a different
stimuli.
The random-dot pattern is used in order to make quantitative statements
about the selectivity of the eye movements and the information they are con-
taining. For this reason the stimuli changes for each search task (2 or 3 colors)
but the search pattern is kept. Furthermore, by repeating the search task with the
same search pattern it can be memorized easily by the subjects. The number of
colors variates between 2 and 3 in order to analyze how the changed complexity
influences the search behavior and the results.
During the whole experiment the eye movement of a subject is recorded.
Later, this data has to be analyzed, and different algorithms are applied. Reports
summarizing the data are generated.
These reports will contain:

• Fixation data
A list of all fixations in the stimuli picture and the surrounding area.


Figure 2.8: Example of target pattern and search picture with 3 colors

• Fixation rankings
A list of the fixations and the surrounding area ordered by either their
frequency or the fixation duration.

• High similarity ranking
For each fixation and it’s surrounding area, a list with patterns that are sim-
ilar to the target pattern can be generated. Similar means e.g. if a fixation
with a 3x3 area in the stimuli picture is given, a similar pattern would be
one where 8 out of 9 squares for each position are the same. A list with all
similar patterns will be created, and the reports present the most frequent
patterns inside this list. This list will be called high similarity list, and the
corresponding ranking high similarity ranking. Also, instead of observing
the frequency of the patterns for building the ranking, the duration of each
fixation can be inherited to each similar pattern - then a ranking where the
similar patterns are ordered by their inherited duration can be built.

• Sub-fixation rankings
For each fixation and its surrounding area, a list of several sub-fixations
can be extracted. These sub-fixations are one square smaller than the target
pattern itself, in a case of a 3x3 target pattern the sub-fixations size is 2x2
squares. A list of the most frequent sub-fixations can be generated. Also,
for these sub-fixations the position inside the fixation area can be used to
build rankings separately for each position.

2.4. EXPERIMENT DESCRIPTION 15

The next chapter discusses in detail the software and algorithm developed
for this work to build, run and analyze this experiment.

Chapter 3

Experiment Software

For this work three different stand-alone applications were developed. These
are:

• Experiment Builder

• Experiment Engine

• Experiment Analysis Tool

The Experiment Builder is an application to create experiment files which
can be run by the Experiment Engine. The Experiment Engine displays the ex-
periment and communicates with the EyeTracker to gather the eye movement
data. The Experiment Analysis Tool creates output files that apply different
analysis algorithms to the experiment data and displays the corresponding re-
sults.
In the following each of these applications and their intercommunication is
discussed in detail.

3.1 Experiment Builder
The Experiment Builder provides the functionality to create different trials for
the experiment within a convenient GUI. With this application, the parameters
for each trial of an experiment is set up and the necessary data files to display
the experiment are created.

3.1.1 Step by Step Experiment Builder
The program is started by running BMPBuilder.exe. It is located on the en-
closed CD in the Folder ”BMPBuilder”. To run properly, it requires Windows

17

18 CHAPTER 3. EXPERIMENT SOFTWARE

Figure 3.1: Startup screen of Experiment Builder

XP and the .NET 2.0 Framework.
In the main GUI that shows up after start of the application, the user begins
with an empty experiment. (Figure 3.1)

Main Window

In the main window, the user has the following choices:

• Adding one ore more trials (a so called block of trials) to the experiment.
A block of trials contains trials with the same set of parameter. When
adding a block of trials, a new window will open and ask the user for
the specific parameters. After entering and confirming these parameters,
the trial block will appear in the main GUI under ”List of trial blocks”.

• Moving up or down a block of trials will change the sequence in which the
experiment is run.

• Deleting of an previously added block of trials.

• Creating of necessary data files for the Experiment Engine.

• Via pull-down menu ”File” an existing experiment can be loaded, a previ-
ously created one saved or a new one started.

3.1. EXPERIMENT BUILDER 19

Figure 3.2: Window to enter trial block parameter

.

Adding a Block of Trials

By clicking on the ”Plus” symbol in the right half of the main window (this
button is always active), a new window will open and prompt the user to enter
different parameters (figure 3.2). The opened window is a modal dialog and will
not let the user return to the main window until either it is closed (which will
not add any trials to the experiment) or until all necessary data fields are filled.
In order to create a new block of trials, the following information must be
provided by the user (figure 3.3):

• Number of trials
Number of trials to which parameters are applied.

• Name of trial block
A name to identify this block of trials.


• Max duration in s
The maximal duration of each trial in seconds before it is timed out - the
”solution” of the trial is displayed and the next trial starts.

• Display X/Y (Text fields disabled)
Defines the resolution of the screen for displaying the experiment. This
value is fixed with 1280x1024 pixel.

• Color definitions
Each color for the squares in the random-dot pictures and patterns can
be defined. The first two colors are always activated. By checking one
or more of the check boxes additional colors can be added. Clicking on a
color itself will display a color dialog where the user can change the color
for the specific selection. Not only the colors for the squares can be set, also
the color for the ”solution” frame (a frame that shows where the pattern
is located in the picture, displayed at the end of a trial), the color for the
surrounding box of the picture (with the displayed border size) and the
background color can be specified.

• Square definitions
In the ”Size X*Y of Square” text fields, the user can enter the width and
height in pixel for each square. If the ”Fill Screen with Squares” check box
is activated (which is in the current program the only possible configura-
tion), the whole screen is used and the total amount of squares for the pic-
ture is calculated and displayed in the ”# Squares X/Y” text fields. (There-
fore, the number of squares in these text fields cannot be modified.) If the
”Quadratic” check box is activated, the picture’s overall shape is quadratic,
otherwise it will have the rectangular shape of the screen. The number
of squares of which the pattern itself consists is entered in the ”Pattern
squares” text fields. They define the number of squares in the pattern - e.g.,
the values 3 * 3 would define a pattern with each three squares in horizon-
tal and vertical direction.

• Pattern definitions
The user can specify if the patterns and pictures are either displayed on
the same screen or on subsequent screens. If checked, the text fields below
are disabled and will not need data input. If not, these text fields define
the time periods the pattern will be displayed before it is replaced by the
picture.

• Preview and cancel/accept
The preview button generates a rough preview of the look-alike of one trial
of the current block. The preview is optimized for the fixed display size
1280x1024. The cancel button will close the trial parameter window and


Figure 3.3: All necessary parameter information is entered

discard all entered information. If all necessary data is entered, the accept
button is enabled and can be used to add the data to the experiment.

After accepting the data input, these parameters are used to create a block of
trials that appears in the trial list of the main window. The total number of trials
is stated below the list of all trials. (”Total number of trials”, figure 3.4)

Additional Data Modifications

Selecting one of the previously added trial blocks allows the user either to delete
it or to move it up or down. The order in which the trials are executed during
the experiment is the same as they are listed in the trial list. Moving up or down
of a trial block will change the execution order. Deleting will remove it from the
experiment. To edit existing experiment data, the user can double-click on the
trial block that needs to be modified.
Via the pull-down menu ”File” it is possible to start over, load or save an ex-
periment (figure 3.5). The user can exit the program at any time. A warning will


Figure 3.4: A new block of trials has been added to the trial list

Figure 3.5: The File menu

appear if there are unsaved changes. The ”?” menu accesses an ”Info” window
which displays contact and program information.

Generating Experiment Data

Finally, when all desired trial blocks are added to the experiment, the user can
generate the experiment ﬁles. Clicking on the ”Generate” button opens a folder
browser dialog where the user can choose the folder to create the experiment
ﬁles in.


3.1.2 Experiment Builder Files
The Experiment Builder will create the following directory structure and the
following files:

• /config.txt
Provides display size and total trial number data in a text file. Will be read
by the Experiment Engine.

• /experiment.bin
Contains the experiment object data which can be read by the Experiment
Builder.

• /data/PatternData X.bin
Contains the pattern object data for trial number X. Is processed by the
Experiment Analysis Tool.

• /data/PictureData X.bin
Contains the picture object data for trial number X. Is processed by the
Experiment Analysis Tool.

• /data/trialData X.txt
Contains text data for each trial X that is needed by the Experiment Engine
and contains e.g. the location of the search pattern in the displayed picture.

• /images/trial X.bmp
The search picture image for each trial X in a bitmap file. The Experiment
Engine will display this search picture during the experiment execution.

• /images/trialPattern X.bmp
If the target pattern and the search picture are not displayed on the same
the screen, this image file containing the search pattern will be displayed
at the beginning of each trial X.

3.1.3 Class Overview
The following section gives a overview of the code structure and the classes
of the Experiment Builder. Additional to the structural overview the complete
code documentation can be found in appendix D. The complete source code, its
project files and its documentation in electronic form are also on the enclosed
CD. (CD-Content is described in appendix A.2)
The Experiment Builder is written in C#, using the .NET 2.0 Framework from
Microsoft. It was developed with Visual Studio 2005. A list of all used software
and external tools is listed in appendix A.1.
The Experiment Builder can be divided into three sub-areas: GUI objects,
experiment data objects and pattern/picture objects.


Figure 3.6: Overview of all classes in the Experiment Builder

The GUI objects represent the user interface, marked by the stereotype <<GUI>>
in the class diagram (ﬁgure 3.6). Its functionality is described in the previous
section 3.1.1. The ExperimentForm object displays the main window and the
TrialForm object is responsible for the trial parameter window. For display-
ing a preview, a new PreviewForm object is created, and to display the info
window the Info object is used.
Constants provides the GUI with already pre-deﬁned data (such as the

3.2. EXPERIMENT ENGINE 25

display size of the experiment screen), mostly provided in static methods. The
GUI itself is invoked by Program, which does nothing else than to instantiate a
new ExperimentForm.
The ExperimentForm and TrialForm user interface objects are using the
ExperimentData and TrialData to store the entered data. An ExperimentForm
starts with an empty ExperimentData object, containing an empty list of
TrialData. When the user adds a new block of trials, an instance of a TrialForm
is displayed. After entering all information into the user inferface object TrialForm,
the data is stored in a TrialData object and then added to the list of trial
blocks in the ExperimentData. TrialData contains also a PreviewData
object, where the PreviewData stores the trial data that is necessary to dis-
play a preview. Hence, the PreviewData is used to show a preview within the
PreviewForm. The ExperimentForm displays the changed list of trials within
its GUI.
The Pattern and Picture objects represent target patterns and search pic-
tures that the Experiment Builder is supposed to generate. They provide func-
tionality to create representative data objects and images of a pattern or picture,
where its data and its corresponding bitmap files can be stored on the hard drive
and later on read and accessed by the Experiment Analysis Tool.
A Picture or Pattern consists of a two-dimensional array of squares, each
square can hold a color value. Because both have many common properties and
methods, both inherit from the class Bitmap. After instantiation of a Picture
or Pattern (using the values the user entered in the GUI) they can be filled with
evenly distributed random color values. After that, a pattern can be connected
with a picture, so that the pattern itself occurs exactly once inside the picture.
When both (picture and pattern) are created and connected, the generation pro-
cess (invoked by the user in the GUI) uses these objects to create and store their
bitmap image objects and their data files (bmp and bin files in the ”data” and
”image” folder).

3.2 Experiment Engine
The Experiment engine displays and runs an experiment created by the Exper-
iment Builder. While running the experiment, it communicates with the Eye-
Tracker device and takes care of controlling the EyeTracker according to the ex-
periment needs.

3.2.1 Step by Step Experiment Engine
Before the Experiment Engine can start, the previously generated files and fold-
ers from the Experiment Builder have to be copied in the same folder where the
executable of the Experiment Engine is located. After starting the Experiment


Figure 3.7: Example for a target pattern

Engine, it will access the file config.txt in the same folder as well as all txt
files in the ”data” Folder and all bmp files in the ”image” folder.
To run the experiment, the executable BMPBuilder.exe must be started.
To run properly, the screen resolution of the host computer should be set to
1280x1024. An existing connection to an EyeLink II eye tracking device from
SR Research is needed to record eye movements, but there is also a simulation
mode in which the Experiment Engine can be run. In the simulation mode the
experiment execution will run without an EyeTracker, but there will obviously
not be any data recording. The compiled program file BMPBuilderSim.exe
enclosed on the CD is compiled in simulation mode. But because of the close
dependency on the hardware and the installed EyeTracker libraries, it is very
likely that the executable has to be recompiled to run.
After execution of the program a brief description of the experiment is given.
Pressing ”Escape” will continue with the calibration screen, in which the cam-
era of the EyeTracker can be focused. Also, the calibration of the gaze fixation
to properly record data has to be adjusted. (The EyeTracker calibration is docu-
mented in [19].) Calibration is not available in simulation mode.
After the calibration each trial will start with the drift correction. The drift
correction tries to correct occasional shift of the headset on the subject’s head.
The subject will look at a dot in the center of the screen and continue to the next
trial by clicking a button on a game pad connected with EyeTracker. (”Escape”
will skip the drift correction in simulation mode.)
After drift correction the target pattern will be displayed to the subject. For
an example of a target pattern see figure 3.7. The subject should memorize this
pattern - the display time depends on the parameters defined in the Experiment
Builder. After the given time period the search picture will appear in which the
subject is supposed to find the target pattern. If pattern and picture are defined
to be on the same screen, it will directly start the trial. See figure 3.9 for an
example screen.
As soon as the subject finds the target pattern in the picture, the trial can
be ended by pressing the button on the game pad (or ”Escape” in simulation
mode). An example for a search picture screen (in which the pattern in figure 3.7
can be found) is shown in figure 3.8. To display the ”solution” of the trial, a red

3.2. EXPERIMENT ENGINE 27

Figure 3.8: The search picture in which the target pattern has to be found

rectangle will mark the position of the target pattern in the picture and therefore
provide feedback to the subject whether he or she was right or wrong. If the
subject does not find the pattern (and does not press the button) the experiment
will timeout after the defined time period. The position of the target pattern will
be marked and the experiment continues with the drift correction for the next
trial.
At the end of the experiment a ”Save File” dialog pops up, and the experi-
menter can choose a folder where the recorded eye data should be stored. The
eye data comes in an edf binary format. A tool provided by SR Research called
Visual2ASC can convert these edf files into ASCII text files (Ending asc). The
edf file will contain samples and events of the eye movement, where a sample
describes each position of the eye for each scan period (e.g. 250Hz). Also the
EyeTracker can calculate fixations, blinks and saccades of the eye and names
them as events. To analyze the recorded tracking data with the Experiment
Analysis Tool, the edf file has to be transformed into an asc file, whereby only
the events are needed.


Figure 3.9: Target pattern and search picture on the same screen

3.2.2 Code Overview
The Experiment Engine is written in C. The source code and the Visual Studio
project files as well as a pre-compiled executable can be found on the enclosed
CD in the folder ”BMPExperiment”. The Experiment Engine uses libraries from
SR Research for communication with the EyeTracker. More information, a pro-
gramming guide and the EyeTracker software is available on the SR Research
Homepage (see [18]). To display graphics on the screen, the Open-Source li-
brary SDL (Simple DirectMedia Layer, see [3]) is used. SDL requires an OpenGL-
compatible graphic card and drivers.
To develop the Experiment Engine, an existing code was used and modi-
fied. This code template called ”picture” is included in the EyeTracker software
packet for developers. It was modified according to the guidelines in the Eye-
Tracker Programming guide (see [17]). The template contains code for display-
ing pictures on the screen, and records the eye movement of the subject. (For
complete documentation of this template see [17, Chp. 17].)
In the following a brief description of the code files is given:

• picture.h
Declarations to link together the template experiment files.

3.3. EXPERIMENT ANALYSIS TOOL 29

• main.c
Main function for windows execution. Setup and shutdown of link to the
EyeTracker device, display of experiment graphics, opens and saves edf
file.

• trials.c
Called to run all trials for an experiment. Performs system setup at the
start of each experiment, then runs the trials. Handles return codes from
trials to allow the subject to end the trial by pressing a particular button.
The code that is changed compared to the template is inside this file.

• trial.c
Implements a trial using the SDL graphics library. Includes the code for
drawing and displaying the experiment picture, for sending messages to
the eye tracker and handling any subject responses.

The experiment first initializes and connects to the EyeTracker. It then creates
a full-screen window, and sends a series of commands to the tracker to configure
its display resolution, eye movement parsing thresholds, and data types. Then
it asks for a file name for the edf file, which it commands the EyeLink tracker
to open on its hard disk.
The program then runs all the experiment trials. Each trial begins by per-
forming a drift correction, recording is then started. After all trials are com-
pleted, the edf file is closed and transferred via the link from the EyeTracker
hard disk to the display PC. At the end of the experiment, the window is erased
and the connection to the EyeTracker is closed.

3.3 Experiment Analysis Tool
The Experiment Analysis Tool takes the data generated during the experiment,
processes them using different algorithms and creates reports . Because there is
no ”standard” way to develop reports showing the algorithm’s results, the Ex-
periment Analysis Tool generates Excel files that make data interpretation easier
for humans.

3.3.1 Step by Step Experiment Analysis Tool
The Experiment Analysis Tool is started by running BMPAnalysis.exe. It is
located on the enclosed CD in the folder ”BMPAnalysis”. To run properly, it
requires Windows XP and the .NET 2.0 Framework. For generating Excel sheets,
Microsoft Office 2003 and the Microsoft Office Primary Interop Assemblies (PIA)
for Office must be installed. (See [10])


Figure 3.10: Experiment Analysis Tool main GUI

In the main GUI that shows up after start of the application, the user begins
with no experiments in the analysis queue. (Figure 3.10)
The next step is to select a folder that contains experiment data. For this pur-
pose the user can click on the folder symbol next to ”Choose Experiment Folder”.
A browse folder dialog will appear, where the desired experiment folder can be
chosen. After selecting a folder, the second folder symbol next to ”Choose Data
File” is enabled - clicking on it let the user choose an asc data file containing the
EyeTracker events from the experiment.
After folder and file are selected, the ”Data File Info” box displays general
information about the experiment, specifically the number of trials, number of
fixations and the average number of fixations per trial. (Figure 3.11)
Now two terms have to be defined that are used within the following expla-
nations.
Region of Interest (ROI):
Each subject’s fixation just points exactly to one square in the pattern. But in or-
der to analyze the fixation, the surrounding squares have to be included, so that


Figure 3.11: The experiment folder and data file have been selected

a region with at least the target pattern size can be used for the analysis. Hence,
the surrounding squares which are needed to reach the target pattern size will
create the fixation used for the analysis. This fixation area will in the following
be called region of interest or ROI. The width and height of the target pattern
builds the size of the ROI. For example, if the size of the target pattern was 3x3
squares, and the fixation was at position X:5 and Y:5 in the picture array, the area
used by the analysis tool would create a ROI with the coordinates X:4 and Y:4 for
the top left corner and X:6 and Y:6 for the bottom right corner. Furthermore, a ra-
dius parameter can be defined. This parameter indicates how much addititional
surrounding squares (additional to the target pattern size) must be included. A
radius of 0 does not increase the ROI. For a radius of 1 the ROI area would be
increased: a fixation at X:5 and Y:5 would lead to a ROI starting with X:3,Y:3 top
left and X:7,Y:7 bottom right. Hence, the value of the radius is added to the used
target pattern size and influences the fixation area used for the calculations.

Fixation pattern of interest (FPOI):
Each ROI used for the analysis can be divided into one or more patterns with


size of the target pattern. E.g. for a given target pattern with 3x3 squares, and a
radius parameter set to 0, the size of the ROI would be 3x3 squares and would
be the same as the size of the target pattern. (In this case the fixation pattern
of interest would be the same as the ROI.) For a radius of 1, the ROI would
have a size of 5x5 squares, and would contain 9 patterns with the same size
as the target pattern. In this case, of each ROI all patterns are extracted which
have the same size as the target pattern. These extracted patterns are used for
further calculations. Each of these sub-patterns extracted from a ROI will be
called fixation pattern of interest or FPOI.

Now the user can define the parameters for the reports that are to be created
by the Experiment Analysis Tool.

• From - To
Defines at which trial number the analysis should start and end.

• Radius
The used radius for creating the ROI.

• Similarity Measure
A percentage value that describes when two FPOIs are meant to be similar.
A FPOI can be compared to another pattern if both have the same size (e.g.
both are 3 squares wide and 3 squares high). The number of matching
colors at each position of both fixations can be expressed as a percent value
- e.g. if both fixations in a 3x3 pattern are the same, but only differ in the
color in their bottom right corner, the similarity of one pattern with the
other would be 0.88 (or 88%). For the algorithm used in the analysis a
list with all possible patterns that are similar to all FPOIs is created. The
generated reports present the most frequent patterns inside this list. This
list is called high similarity list.

• Include Pattern Fixation
If this option is checked, ROIs whose surrounding squares include the tar-
get pattern are processed in the analysis. Otherwise, these ROIs are left
out. By default this is not checked. Analyzing fixations that include the
target pattern could lead to altered results.

• Include 1st Fixations
If this option is checked, the first subject fixation in the picture for each
trial is processed in the analysis. By default, this is not checked. At the
start of each trial the first fixation is based on only a coarse assessment of
the new scene and therefore is not fine-grained enough for the purpose of
the current study.


• Move fixation center
By default, a fixation is supposed to be the center of the ROI. For building
the ROI, the surrounding squares depend on the target pattern size and the
radius. With this option the center of the ROI can be moved by 1 square
to all eight possible directions (left, top left, top, top right, right, bottom
right, bottom, bottom left). E.g., to create a ROI the squares which are
surrounding the center are used. With this parameter, the center can be
moved either horizontal, vertical or diagonal. So if the center is moved to
the top-left, the squares which are right and below the subject’s fixation
are used to built the ROI.

• Display Top
At the end of each trial block, several rankings are displayed (e.g. the
FPOIs ranked by their frequency). A detailed discussion of these rankings
are found in section 3.3.2. This value defines for all rankings how many of
the top ranks are listed in the analysis.

• Short Excel Report
By default, the Excel report sheets display a list with fixation details, in-
cluding their position and a graphical representation of their ROI. Check-
ing this option will leave out the fixation details and print only the rank-
ings.

After all parameters are set (or left to their default values) the experiment
can be added to the analysis queue by clicking the ”Plus” button. Once added,
it can be selected and then removed from the queue with the ”Minus” button.
The desired amount of experiments that should be analyzed can be added to
the queue, by repeating the steps above. After all experiments are in the queue,
the report creation can be started. The user can choose between generating an
Excel sheet or raw data files (which are text files). (Figure 3.12)
The processing of all experiments can take a while, depending on the amount
of trials and colors. In the following the output files are discussed in detail.

3.3.2 Analysis Output Files
The Experiment Analysis Tool can create two different output files: Excel sheets
or raw data (which means text) files. In the following both outputs are described
in detail. Beforehand the term ”Vector” has to be explained, because the follow-
ing explanations are referring to it.

Vector:
A vector is a one-dimensional representation of a (two dimensional) pattern. If
the pattern is e.g. a 3x3 array of different color values, a vector of this pattern


Figure 3.12: A queue was generated and the experiment analysis started

would be an one-dimensional array with the length 9. The first three color val-
ues of the vector would be identical with the top three color values of the pattern
array. (The top left, top center and top right colors.) The second group of three
would correspond to the colors in the middle row, and the the last three values
would be the ones in the bottom row. Therefore, the conversion is row-wise
from top to bottom. A vector can be written as a string by writing the abbrevi-
ated color letters in a single row (e.g. ”WWWB” is a vector string for describing
a vector with the color sequence white,white,white and black).
Furthermore, a vector has the properties ”frequency” and ”weight”. If a
set of vectors are combined in a list, the frequency describes the amount of oc-
currences in this list, and the weight any assigned value. This list can now be
either sorted by frequency or weight. (In the following the value assigned for
the weight is always the duration of the subject’s fixation.)
It is possible to calculate the average vector of any number of given vectors.
To calculate an average vector, first the frequency of each color at the first po-
sition in every vector is counted. Then the color with the highest amount of


Figure 3.13: Experiment Metadata

occurrences is written at the first position of the average vector. This will be
repeated for every position.
An example for building the average vector: Three vectors with two dif-
ferent colors and a pattern size of 2x2 squares are given. The first vector is
”WWWB”, the second ”WWBB” and the third ”WBBB”. The average vector
would be ”WWBB”. For the first position the frequency for the color ”W” (white)
is 3, and for ”B” (black) 0. For the second position it is for ”W” 2 and for ”B” 1,
for the third position ”W” 1 and ”B” 2 and for the forth position ”W” 0 and ”B”
3. Hence, the color with the highest frequency is taken into the average vector.
Also the average vector can be built using the weight: Now not the frequency
decides which color is taken, but the color with the highest sum of weight de-
fines the new average vector. E.g., if there are four vectors ”BBWW”, ”BWWW”,
”BWWB” and ”WBBB” with their corresponding weights 50,75,100 and 200, the
average vector would be ”BBWB”.
If there is no color with the highest frequency or the highest weight (because
two or more colors have an equal value), then the color which has the highest
value in a given color distribution is taken. The color distribution is a parameter
which is always provided when calculating the average vector, normally this is
the overall statistical average color distribution of all FPOIs. E.g., for two colors
a color distribution could be 57% for white and 43% for black.
Two vectors can be compared to each other. The result of such a compari-
son is a percentage value that indicates how many positions in the two vectors
contain matching colors. Comparing the vectors ”WWBB” and ”BWBW” would
return the value 0.50 or 50%.

Excel Output File

The Experiment Analysis Tool will generate a new Excel sheet for each new
block in the experiment. In the following the content of such an Excel sheet is
discussed. Examples for complete results are found in the ”BMPExperiment-
Data” folder on the enclosed CD.
The first two lines with the headline ”Experiment Metadata” locate the data
file and experiment folder of the output. The subsequent line describes the pa-
rameter set of the given output. (Figure 3.13)
Now a detailed output of data for each trial is given. (This output is only
displayed if the Short Excel Report parameter is not set.) At the beginning of each


Figure 3.14: Trial information

Figure 3.15: Details for four different fixation types, including ROI and target
pattern

trial, the trial number, name of the trial, block number the trial is part of, the
information whether this trial was timed-out or ended by the subject via button
press and overall duration of the trial in ms is displayed. (Figure 3.14)
For each trial, the ROI (region of interest, an area surrounding the fixation
and used for the analysis) and additional details of a subject’s fixation is printed.
The size of the ROI depends on the used radius parameter. The ROI can be split
up in different patterns with the size of the target pattern, the FPOIs (fixation
patterns of interest, which are patterns with the size of the target pattern and
are extracted from the ROI). For a radius of 0, this would be equal to the printed
ROI. For a higher radius the number of FPOIs is higher.
These details include: (see also figure 3.15)

• Type
Has one of these values:

– ”Pattern” to identify this not as a fixation but as the description of the
target pattern.
– ”Fix” for a normal fixation inside the picture.
– ”FIX(1st)” to identify this fixation as the first fixation of the trial.
– ”PFIX” indicates that the ROI includes the target pattern.
– ”OUTSIDE” for a fixation on the screen outside the picture.

• PosX/PosY
Describes at which square in the search picture array the subject’s fixation


was in horizontal/vertical direction, starting with 0. The maximum value
is the size of the picture array in horizontal/vertical direction minus one.

• SquaresX/SquaresY
The width and height of the ROI, with the fixation position used as center
(if not differently defined in the parameters by the MoveX/MoveY fields).
This is the target pattern width/height when a radius of 0 is used and
more if the radius is higher. In the current state of the Analysis Tool only
quadratic ROIs are used, therefore width and height of this rectangle is
always equal.

• ViewX/ViewY
At the end of the line the ROI is printed. The values for ViewX/ViewY are
coordinates that describe the position of the top left corner of the ROI in
the picture.

• Dur.
The duration of the subject’s fixation in ms.

• Pup.Size
The subject’s pupil size in pixel. The value strongly depends on the dis-
tance between the camera and the subject’s eye. The pupil size can de-
scribe the attention - if the subject pays a lot of attention to the current
trial, the pupil size will get bigger, if not, it will get smaller. (See [6])

• Distr.
The color distribution of all colors in the current ROI. It displays an abbre-
viation for the color and a percentage of the color distribution.

• Avg. Vector
This vector string describes the calculated average vector of all FPOIs that
can be extracted from the current ROI. If the ROI has the same size as the
target pattern (that means the radius parameter is 0), the average vector is
equal to the ROI.

• Vector Comp.
The vector comparison value is a percentage which describes how many
positions in the average vector are matching the target pattern.

• Printed Pattern
A graphical representation of either the target pattern or the ROI is printed
at the end of each line.

At the end of the list that describes the ROI and fixation details for a trial
block (where each Excel sheet contains exactly one trial block) an analysis with
color distributions data and different rankings follows.


Figure 3.16: Color distribution example

Figure 3.17: Average vector of all FPOIs

First of all, the parameter set is printed again. Then the overall statistical
color distribution for all ROIs is given. After this comes a detailed list of the
color distribution, separated into bins and ordered by their frequency, related to
the size of the target pattern. This is because the color distribution has always
discrete values, e.g. for a 2x2 target pattern, the color distribution is either 0%,
25%, 75% or 100% for each color. The distribution list ranks the frequency of
each color for each possible percent value, related to the pattern size and the
number of colors. (Figure 3.16)
In the next part of the rankings the average vector is built, combining all
FPOIs which are extracted out of the ROIs for the analysis. The average vector
is displayed as a vector string (the abbreviated color letters in a single row),
followed by its color distribution and a graphical representation of this vector.
(Figure 3.17)
Next there are several ”Top-X”-Rankings. X is a number defined in the pa-
rameter set and describes how many of the first patterns in the rankings are
displayed. It starts with a ranking that lists the top X of all FPOIs, and it is called
”fixation ranking”. All FPOIs are here ranked by their frequency. (Figure 3.18)
Each ranking starts with a repetition of the target pattern. Then the ranks
for the first places are displayed, including the vector string of each pattern


Figure 3.18: Top 3 of a fixation ranking

in the ranking, as well as the color distribution of this pattern, the amount of
occurrences (frequency) in the ranking, the calculated weight and a graphical
representation of this pattern. In all rankings the weight is the total time of the
fixation duration, inherited from the ROI from which the FPOI was extracted.
The similarity is a percent value and describes how similar the described pat-
tern is to the target pattern.
After the third and each fifth place an average vector of all precedent ranking
vectors is built and displayed. This means that all previous patterns and their
vectors are combined to an average vector by counting the frequency of each
color for each position in these previous patterns, and keeping the color with the
highest frequency at each position in the vector. If the frequency for a position is
the same for all colors, the color which has the highest value in the overall color
distribution is taken. An average vector using weight means that the weight
of each vector is added up for each color instead of the frequency - the color
with the highest weight is taken into the average vector. (Figure 3.18 shows an
example for the average vector in a ranking.)
The same ranking is generated for all possible patterns that have a ”high
similarity to the target pattern.” That means that for each FPOI all patterns are
calculated which are similar to this FPOI, depending on the defined similarity
parameter. Then the patterns are added to the high similarity list and the same
procedure is repeated with the next FPOI. E.g., if the similarity value is 0.80,
and the target pattern size 3x3, the high similarity list would contain all patterns
which are equal to each FPOI but differ in only one color. Therefore, the ranking


Figure 3.19: Example for the high similarity ranking

displays the most frequent patterns in the high similarity list. In difference to
the fixation ranking this ranking can contain the target pattern. (Figure 3.19)
Also a ranking by a defined score system is created. This ranking takes every
pattern inside the high similar list and compares it again with each possible
pattern. If the similarity is higher than 60% (a fixed value), the pattern will get
a score, depending on the similarity to the possible pattern. E.g., if both of them
have in 7 out of 9 positions the same color, the similarity value would be 77,
and this value would be added to the score of the pattern. This ranking was
implemented to see if it would give interesting results, but did not fulfill the
expectations and is not regarded in further discussions.
The next ranking shows a placement of the ”sub-fixations”. A FPOI is split
up into 4 smaller patterns, which are in the following called ”sub-fixations”. A
sub-fixation is one square smaller in width and height than the FPOI. E.g., given
a FPOI of 3x3 squares, the size of a sub-fixation would be 2x2. Each one of the
four sub-fixations starts at a different corner of the FPOI (top left, bottom left, top
right, bottom right) whereby the sub-fixations overlap. The sub-fixation lists can
now be ranked by their frequency. There are five rankings for the sub-fixations:
four for the different corner positions, and one where all sub-fixations are not
separated by their corner position. Figure 3.20 shows an example for the overall
sub-fixation ranking.
The next ranking is an approach that tries to combine the sub-fixations for
each corner position into a new ranking. Place one in this ranking is created by
taking the top-position of each corner, and putting them together into a pattern


Figure 3.20: Example for the overall sub-fixation ranking

with the target pattern size. At each position where a sub-fixation overlaps with
another sub-fixation, the sub-fixation with the highest frequency value is ”on
top”. This approach tried to combine the sub-fixations into possible patterns
that would be close to the target patterns, but did also not meet the expectations
and is not regarded in further discussion.

After this the same rankings are repeated, but this time not ordered by fre-
quency but by its weight - which is the duration of each subject’s fixation.

Raw Data Output Files

The raw data files require a folder, which the user specifies when he clicks the
generate button for the raw data. Inside this folder, for each block of trials three
different files are created. For each file, the X marks the block number, starting
with 1.

• lookupX.txt
A so called look-up file, which creates all possible patterns, depending
on the height, width and number of colors of the target pattern of a trial
block. For each possible pattern an ID is given, starting with 1 for the first
possible pattern, counting up. This way each possible pattern that can be
created has an unique ID.

• rawDataX.txt
For each FPOI, the previously created ID of this pattern is written in each
line, followed by the color distribution for each color. At the beginning of
each block, a java style comment with the parameter set is printed. At the


beginning of each trial, a java style comment with the target pattern ID is
printed.

• resultX.txt
This file contains a text style version of the fixation and high similarity
rankings as described in the Excel output.

The goal of the raw data files was to use a Data Mining Tool such as kNime
(see [8]) to analyze the data in an easy way. Unfortunately, the data analysis with
kNime did not lead to valuable results and is not regarded in further discussion.

3.3.3 Class Overview
In the following a brief overview of the classes used for the Experiment Analysis
Tool is given. The Experiment Analysis Tool consists of two parts: A class library
that represents the different eye events of the generated asc data file from the
EyeTracker, called ”ASCLibrary”. The other part contains classes for the GUI
and the generation of the output files from the Experiment Analysis Tool and is
called ”BMPAnalysis”.
The following section gives a rough overview of the code structure and its
classes. Additional to the structural overview the complete code documentation
can be found in appendix C for the Experiment Analysis Tool and appendix B for
the ASCLibrary . The complete source code, the project files and its documenta-
tion in electronic form is located on the enclosed CD in the folders ”ASCLibrary”
and ”BMPAnalysis”.
The Experiment Analysis Tool is written in C#, using the .NET 2.0 Frame-
work from Microsoft ([10]). It was developed with Visual Studio 2005. A list of
all used software tools is listed in appendix A.

ASCLibrary Class Overview

The ASCLibrary is a representation of an EyeTracker data file. It is a library
that provides the possibility to access the eye tracker data and events stored
in the asc file. Each asc file contains data information for each trial run dur-
ing the experiment. The ASCData object represents a list of trials of the type
ASCTrialData, and provides a method to load all trials from the file into this
object. Each trial is therefore represented by an ASCTrialData object that con-
tains its meta data and a list of all fixations, saccades and blinks of the subject’s
eye. The fixations, saccades and blinks are called events. Each fixation is stored
in an ASCFixation object, containing start and end time in ms relative to the
start of the experiment, the duration, the average pixel coordinates of the fixa-
tion on the screen and the average pupil size in pixel during the fixation. A sac-
cade and its ASCSaccation object describe the start and end time of a saccade,


Figure 3.21: Class diagram of ASCLibrary

its duration, the start and end coordinates on the screen, the saccadic amplitude
and the peak velocity of a saccade. (The made-up word ”Saccation” was used
unintendedly.) A blink event and its ASCBlink object represent a blink of the
eye and contain the start and end time as well as its duration of the blink. The
correlation of the objects is shown in the class diagram in figure 3.21.

Experiment Analysis Tool Class Overview

The Experiment Analysis Tool consists of objects for creating the GUI, namely
MainForm and Info (figure 3.22). Program simply invokes the main window
MainForm, and Info displays version and author information. The MainForm
object represents all functions as described in section 3.3.1. The entered infor-
mation, specifically the list of experiments to be analyzed and its parameters,
folder and data file, are stored in an AnalysisObject. This object contains
the list of parameter in its property OutputFileParameter, and some spe-


cific experiment data in the Config object (this is the representation for the file
config.txt in the experiment folder). Furthermore, the analysis objects use
the existing Bitmap object from the Experiment Builder (see 3.1.3) to load and
access the experiment data, as well as the ASCLibrary to load and access the
data stored in the eye tracker data file.
A list of the current analysis objects in the queue is stored and displayed by
the GUI. When the user starts one of the file generation processes (either Excel
sheets or raw data), the appropriate child of the OutputFile object is gener-
ated and fed with the necessary parameters and experiment data. OutputFile
is an abstract object - depending on the type of output desired, either an
OutputFileExcel or OutputFileRawData object is instantiated. OutputFile
implements all necessary calculations for the analysis of the experiment data -
they are the same for both output types. For the calculations several Vector ob-
jects are used, which represent patterns and are handy to do calculations within
these patterns. (A description of a vector is given in 3.3.2.) Also the ColorDistribution
object is used to help with color distribution calculations. OutputFile pro-
vides so called hook-up methods which are used by either OutputFileExcel
or OutputFileRawData to display the calculation results.
Hence, the calculation algorithms are implemented in the OutputFile ob-
ject. The result of these algorithms are displayed within the rankings in the
output files created with OutputFileExcel or OutputFileRawData.
A discussion of a conducted experiment and its results is given in the next
chapter.


Figure 3.22: Class diagram of the Experiment Analysis Tool

Chapter 4

Experiment Setup, Run and
Analysis

4.1 Experiment Setup
4.1.1 Participants
This research was carried out with the assistance of 14 participants of ages 19 to
36, divided into two groups for different experiment setups. Participants were
paid a $10 honorarium. One of the participants was the author, two were lab
members. The rest was naive as to the purpose of the experiment. All of them
had either normal or corrected-to-normal vision. There were two different ex-
periment setups - 5 participated in the ﬁrst setup, 9 in the second.

4.1.2 Apparatus
Stimuli were presented on a DTI Virtual Window 19” TFT-Display. The reso-
lution of this display was set to 1280x1024 pixels and its refresh rate to 60 Hz.
Participants sat approximately 60 cm from the screen. The horizontal and verti-
cal viewing angles of the stimuli were approximately 34 degrees and 26 degrees,
respectively. An SR Research EyeLink II system was used to track eye move-
ments. The average error of visual angle in this system is 0.5 degrees, and its
sampling frequency was set to 250 Hz. A game-pad was used to register the
participants’ manual responses.

4.1.3 Procedure
The experimenter began each experiment by providing the participant with task
instructions, ﬁtting the eye tracking headset to the participant, and then cal-

47

48 CHAPTER 4. EXPERIMENT SETUP, RUN AND ANALYSIS

ibrating the EyeTracker. Participants started each trial, and at the same time
performed a drift correction of the headset, by pressing a button while fixating
on a central marker. Each trial began with the presentation of a target pattern
for about 6 seconds, during which the participants were to memorize the target
structure. After three trials, the display time was reduced to 3 seconds, and after
15 trials to about 2 seconds. Subsequently, the search picture appeared. Partici-
pants were to search the picture array for the target pattern and, while fixating
on the pattern, to press a button on a game pad to terminate the trial. If a partic-
ipant did not press the button within a specific time interval after search display
onset, the trial would time out and the experiment would continue with the next
trial. After termination of each trial, a red box framed the actual target pattern
position, so that feedback to the participants about their accuracy was provided.
The pattern and the picture array consisted of squares in either two or three
colors combined to a random-dot array. The distribution of the squares was
random, while for 2 colors the possibility of appearance for each color was 50%
and for 3 colors 33% each. (See figure 2.7 and figure 2.8 for examples.)
There were two different experiments: The first included 4 blocks of trials,
the second 2 blocks of trials. In both each block contained 30 trials. For the first
experiment, in the first two blocks the size of the squares were 60x60 pixel, and
in the last two 40x40 pixel. The first and third block had squares in 2 colors, and
the second and last block in 3 colors. For the second experiment both blocks
contained squares with a size of 25x25 pixel. The first block was with 2 colors,
the second block with 3 colors. The background color for both experiments was
always white.
For each block of trials, always the same pattern was used. This means that
a subject had only to memorize 4 different patterns for experiment I (two with
2 colors and two with 3 colors) and only two different patterns for experiment
II (one with 2 colors and one with 3 colors). In order not to bore the subject, the
display time of a pattern was longer for the first three trials (6 seconds), and then
reduced (3 seconds). After 15 trials it was reduced again (2 seconds), because the
subjects now for sure memorized the pattern. By using the same pattern for a
block of trial, the possibility that a subject would not be able to memorize the
target pattern was excluded.
It follows a detailed parameter list for both experiment setups:
Experiment I:

• Trial block 1 ”2 Colors big”
Square size 60x60 pixel, 2 colors, 30 trials, array size: 3x3 for target pattern,
17x17 for search picture, maximum duration 35 sec.


4.2. EXPERIMENT RUNS 49


• Trial block 3 ”2 Colors small”

• Trial block 4 ”3 Colors small”

Experiment II:



Both experiment setups are on the enclosed CD. (”BMPExperimentData/
template1.bin” and ”BMPExperimentData/template2.bin”, to be opened with
the Experiment Builder, see section 3.1.)

4.2 Experiment Runs
The first experiment was run with 5 subjects. An experimenter could see the eye
movements of the subject on the eye tracker display screen. The EyeTracker data
files were saved and later analyzed with the Experiment Analysis Tool (See sec-
tion 3.3). After a first brief analysis of the results and observation of the subjects
during a trial it was discovered that the number of fixations for each block of
trials was relatively low, according to the ”big” square size and the easiness for
a subject to find the target pattern (subjects were able to find the target pattern
very quickly). Also, they just started to scan the screen in a systematically way,
e.g. line by line in reading direction. Therefore the overall fixations could have
consisted of a more or less evenly distributed list of fixations on the screen.
The smaller the size of the squares, the more subjects tended to ”jump” with
their fixation randomly across the search picture - it seemed that for a smaller
square size the subjects scanned ”unsystematically” for the pattern. Another
observation for experiments with a smaller square size was that the pattern was
harder to find and more fixations were needed to find it - also it could be ob-
served that the subjects looked more often at already processed parts of the im-
age, which was a sign for an unsystematical scanning for the target pattern.


Because of this the second experiment was run with 9 subjects, in the as-
sumption it would deliver more fixations and more ”intuitive” eye movements
of the subject. Though the results could be more inaccurate because of the Eye-
Tracker inaccuracy, the second experiment had a higher number of subjects’
fixation and seem to provide more useful empiric data for the analysis of the
experiment.

4.3 Experiment Analysis
For each subject, excel files were created with the Experiment Analysis Tool. (See
section 3.3.1) The following parameters were applied:

• Similarity measure: 0.85

• Radius 0 and 1

• All trials

• Pattern fixations not included

• First fixations not included

• Fixation center is in the center of the pattern

A list of all created reports that contain the analysis for experiment I and
II are on the enclosed CD in the subfolders of ”BMPExperimentData”, where
every folder starts with the prefix ”exp NAME” for the first experiment and
”exp2 NAME” for the second experiment. The Excel reports are located under
”xls”. The filename contains the subject’s name, the block number and the used
radius. The short versions of these reports include the word ”short”.
The reports were generated for a radius of 1 and 0, therefore there are statis-
tics for each radius.

4.3.1 Reports Experiment I
The target patterns for each trial block of the first experiment are named with the
letters A to E for each of the five subjects and 1 to 4 for each trial block. Figure
4.1 shows the patterns A1-A4, figure 4.2 B1-B4, figure 4.3 C1-C4, figure 4.4 D1-D4
and figure 4.5 E1-E4. The size in this pattern figures is relative to the size during
experiment display. However, in further figures the patterns are displayed in a
unique size.
Figure 4.6 shows an example of the recorded fixations for subject E (trial no.
67) superimposed in green. The target pattern E3 is marked in red.

4.3. EXPERIMENT ANALYSIS 51

Figure 4.1: Patterns A1 to A4

Figure 4.2: Patterns B1 to B4

Figure 4.3: Patterns C1 to C4

Figure 4.4: Patterns D1 to D4

Figure 4.5: Patterns E1 to E4


Figure 4.6: Fixations superimposed in green, target pattern E3 in red.


Average Vector and Color Distribution

The average vector of all FPOIs (a pattern that describes the most frequent color
for each position in all FPOIs) was build for each trial block. Also the overall
color distribution was calculated.
For the trials with 2 colors the average vector was in nearly all cases a unique
white or black pattern. Whether it was black or white depended on the stronger
value in the average color distribution. If the target pattern was more white
than black (that means, 5 or more out of the 9 squares in the target pattern were
white) the average vector was in most cases more white than black.
Out of this perception, the question arose if the color distribution of the tar-
get pattern can be associated with the color distribution of the FPOIs. Table 4.1
shows the color distribution of the target pattern and the statistical average color
distribution of all FPOIs for block 1, where a radius R of 1 or 0 was used for the
calculations. Table 4.2 displays the same overview for block 3.

Table 4.1: Color distribution block 1 (Exp. I)

Pattern Target Pattern Distr. FPOIs R1 FPOIs R0
A1 78%/22% 57%/43% 62%/38%
B1 56%/44% 54%/44% 57%/43%
C1 44%/56% 47%/53% 45%/55%
D1 33%/67% 45%/55% 44%/56%
E1 67%/33% 56%/44% 58%/42%
Values are white/black percentages

Table 4.2: Color distribution block 3 (Exp.I)

Pattern Target Pattern Distr. FPOIs R1 FPOIs R0
A3 33%/67% 43%/57% 40%/60%
B3 44%/56% 51%/49% 51%/49%
C3 89%/11% 55%/45% 57%/43%
D3 67%/33% 54%/46% 55%/45%
E3 44%/56% 53%/47% 53%/47%

Regarding the color distribution for block 1 and 3, a direct connection be-
tween the color distribution of the target pattern and the color distribution of
all FPOIs can be made - if white is the stronger color in the target pattern, the
overall color distribution of all FPOIs for white is higher, and vice versa.
Regarding trial blocks 2 and 4 which used 3 colors, the average pattern of all
FPOIs mostly contained patterns dominated by black - it seems that black was
a dominant color. The color distribution of the target pattern and the statistical
average color distribution of the FPOIs are presented in table 4.3 for trial block


2 and in table 4.4 for trial block 4. They show that black has always the highest
percentage.

Table 4.3: Color distribution block 2 (Exp. I)

Pattern Target Pattern Distr. FPOI R1 FPOI R0
A2 33%/33%/33% 29%/38%/32% 28%/39%/33%
B2 33%/33%/33% 30%/35%/34% 29%/36%/34%
C2 44%/22%/33% 32%/35%/32% 32%/35%/32%
D2 22%/44%/33% 31%/40%/29% 31%/41%/27%
E2 22%/33%/44% 31%/36%/33% 30%/36%/34%
Values are white/black/gray percentages

Table 4.4: Color distribution block 4 (Exp.I)

A4 33%/56%/11% 30%/40%/30% 28%/42%/28%
B4 44%/44%/11% 34%/36%/29% 34%/38%/28%
C4 33%/56%/11% 31%/38%/30% 31%/39%/29%
D4 22%/67%/11% 29%/42%/28% 28%/46%/26%
E4 44%/33%/22% 35%/34%/31% 35%/34%/30%


Fixation and High Similarity Rankings

For each FPOI a list with patterns that are similar were built. A pattern was
defined as similar to a target pattern when the similarity was higher than 85%,
which means that 8 of 9 colors at each position where in both patterns the same.
A ranking sorted by the frequency of all FPOIs (called fixation ranking) and
their similar patterns (called high similarity ranking) were created with the Ex-
periment Analysis Tool. For a detailed discussion of these rankings see section
3.3.2.
The number of possible patterns that exist for a size of 3x3 squares and 2
colors is 512. For 3 colors it is 19683. This means that the list with similar patterns
may contain up to 512 patterns for 2 colors and up to 19683 patterns for 3 colors.
To display all generated rankings would exceed the scope of this work. In
the following comprehensive tables are presented that attempt to summarize the
results of the high similarity rankings.
Table 4.5 shows the results of the high similarity list for trial block 1, with a
radius R of 1 or 0. The first column names the target pattern, the second and
forth column the rank of the target pattern in the high similarity rankings (for
different radii), and the third and fifth column the similarity of the average pat-
tern to the target pattern (also for different radii). The average pattern was build
out of the first three places. A value of 100% means it is the same as the target
pattern, 88% means that 8 of 9 positions have the same color, and so on. With
the average pattern the approach was to combine the most similar patterns in
such a way that they would result in the target pattern. The average pattern was
build for the first three, five and ten places of a ranking - the average patterns
for the first three places showed the best results, therefore the similarity of this
pattern to the target pattern is printed in the results. It is a good indicator how
similar the top patterns in a ranking have been - a similarity of 77% (7 out of 9)
or above would indicate three highly similar patterns in the top three places of
the high similarity ranking.

Table 4.5: Target pattern rankings in high similarity list, block 1

Pattern Ranking R1 Sim. AvgP. R1 Ranking R0 Sim. AvgP. R0
A1 Rank 4 66% Rank 1 88%
B1 Rank 2 100% Rank 50 33%
C1 Rank 76 33% Rank 239 55%
D1 Rank 2 88% Rank 13 55%
E1 Rank 5 55% Rank 1 100%

A graphical example for a fixation ranking and the corresponding high sim-
ilarity ranking for the target pattern D1 is displayed in figure 4.7. The fixation
ranking (left side) contains the rank for each pattern, the frequency of this pat-


Figure 4.7: Pattern D1 fixation and high similarity ranking, radius 0

tern in the list, the weight (which is the summed duration of all fixations of this
pattern), its similarity to the target pattern and the pattern itself. The high sim-
ilarity list (right side) differs in a way that it contains up to eight times more
patterns. The high similarity list may contain the target pattern, the fixation list
not.
Table 4.6 shows a summary for trial block 3, which had like trial block 1 only
two colors, but used a smaller square size. Additional to this table figure 4.8
depicts the high similarity ranking of the target pattern D3 (used radius 0) and
figure 4.9 pattern C3 with radius 1.
Table 4.7 and table 4.8 show a summary for trial blocks 2 and 4, which both
have 3 colors and differ in their square size. The algorithm seems to work poorly
for 3 colors, compared to the results for 2 colors. But the fixation and high sim-
ilarity ranking for target pattern E4 (figure 4.10) is a good example that in some



A3 Rank 12 66% Rank 18 66%
B3 Rank 61 44% Rank 105 44%
C3 Rank 3 88% Rank 4 77%
D3 Rank 8 77% Rank 1 100%
E3 Rank 6 66% Rank 42 77%

Figure 4.8: Pattern D3 ﬁxation and high similarity ranking, radius 0

cases the basic shape is rotated or moved into one direction and can be identi-
ﬁed.


Figure 4.9: Pattern C3 ﬁxation and high similarity ranking, radius 1


A2 Rank 14 77% Rank 301 55%
B2 Rank 956 33% Rank 1111 11%
C2 Rank 1159 22% Rank 2327 33%
D2 Rank 4683 44% Rank 10225 33%
E2 Rank 176 22% Rank 1413 11%


A4 Rank 52 66% Rank 259 44%
B4 Rank 457 77% Rank 203 44%
C4 Rank 5163 33% Rank 899 55%
D4 Rank 886 55% Rank 1965 66%
E4 Rank 580 11% Rank 412 0%


Figure 4.10: Pattern E4 ﬁxation and high similarity ranking, radius 1


Figure 4.11: Sub-fixation ranking for pattern A4, radius 1

Sub-fixations

Next, an analysis of the rankings for the sub-fixations was conducted. A sub-
fixation splits up each FPOI into four 2x2 sub-patterns. E.g., the FPOIs have a
size of 3x3 squares, so there are four 2x2 fixations inside this FPOI. A list contain-
ing all sub-fixations for each FPOI can be created and ranked by the frequency of
a sub-pattern in the list (called sub-fixation ranking). The sub-fixation ranking
shows the most frequent sub-fixations ranked by their frequency of occurrences
in this list. Figure 4.11 shows the Top 10 of sub-fixations for pattern A4 (used
radius 1).
It is possible to separate the sub-fixations into 4 different sub-fixation rank-
ings, each one for their position in the FPOI: a ranking for all sub-fixations of the
top left corner, of the bottom left corner, of the top right corner and of the bot-
tom right corner. But if done so, the generated list produces similar results for
each corner. Compared to a list where they are not separated by their position


in the FPOI there is no big difference between them. Therefore, there is no extra
information by separating the sub-fixations by their positions.
Table 4.9 shows an overview for trial block 1 and 3 (both trial blocks with 2
colors, differing in their square size) in which it is checked if the first sub-fixation
in the sub-fixation ranking was part of the target pattern. If it was not part of
it, the rank is printed at which the first sub-fixations in the rankings occurs that
is part of the target pattern. Table 4.10 does the same for trial block 2 and 4 (3
colors, different square size).

Table 4.9: First place in sub-fixation rankings is part of target pattern, block 1
and 3
Pattern Block 1 R1 Block 1 R0 Pattern Block 3 R1 Block 3 R0
A1 No (Rank 2) Yes A3 Yes Yes
B1 Yes Yes B3 Yes Yes
C1 Yes Yes C3 Yes Yes
D1 Yes Yes D3 Yes Yes
E1 No(Rank 2) Yes E3 Yes Yes

and 4
Pattern Block 2 R1 Block 2 R0 Pattern Block 4 R1 Block 4 R0
A1 Yes Yes A4 Yes Yes
B1 No (Rank 2) Yes B4 Yes Yes
C1 Yes Yes C4 Yes Yes
D1 Yes Yes D4 No (Rank 2) Yes
E1 Yes Yes E4 Yes Yes

The results show that a remarkably high number of sub-fixations on first
place in the rankings were part of the target pattern. The overall number of
possible sub-fixations with a size of 2x2 squares is 16 for 2 colors and 81 for 3
colors. Especially for 3 colors the results are remarkably good.


Duration of Fixations

One way to rank the fixation and high similarity lists is to sort them by their fre-
quency - another is to use the fixation duration which can be used as a weighting
factor for a ROI and the extracted FPOIs. (See section 3.3.2 for details about using
the fixation duration as weight for a ranking.) An examination of rankings that use
the duration of the eye fixation as a weight and sorting factor for the ranking
instead of the frequency is done in table 4.11 for trial block 1, table 4.12 for trial
block 2, table 4.13 for trial block 3 and table 4.14 for trial block 4. The first column
names the pattern, the second and forth the rank in the high similarity ranking,
which is ordered now by the overall duration of the eye fixation. The third and
fifth column give the similarity of the average vector of the first three places to
the target pattern. The values in brackets are the already discussed values for
the high similarity ranking ordered by the frequency. They are printed again to
make it easier to compare both high similarity rankings.

Table 4.11: Target pattern rankings in high similarity list using duration as
weight, block 1

A1 Rank 2 (4) 100%(66%) Rank 23 (1) 33%(88%)
B1 Rank 2 (2) 88%(100%) Rank 25 (50) 66%(33%)
C1 Rank 103 (76) 44%(33%) Rank 352 (239) 44%(55%)
D1 Rank 112 (2) 77%(88%) Rank 6 (13) 55%(55%)
E1 Rank 63 (5) 55%(55%) Rank 88 (1) 88%(100%)

weight, block 2

A2 Rank 86 (14) 55%(77%) Rank 2 (301) 88%(55%)
B2 Rank 1005 (956) 22%(33%) Rank 969 (1111) 22%(11%)
C2 Rank 1361 (1159) 44%(22%) Rank 3192(2327) 11%(33%)
D2 Rank 6004 (4683) 44%(44%) Rank 10265(10225) 44%(33%)
E2 Rank 1292 (176) 66%(22%) Rank 1430 (1413) 11%(11%)

Including the duration as a sorting factor for the rankings showed no im-
provements. Also the rankings by duration seem to be overall worse than the
rankings by frequency.


weight, block 3

A3 Rank 87 (12) 55%(66%) Rank 32 (18) 66%(66%)
B3 Rank 59 (61) 44%(44%) Rank 92(105) 33%(44%)
C3 Rank 69 (3) 77%(88%) Rank 154(4) 77%(77%)
D3 Rank 42 (8) 33%(77%) Rank 24 (1) 44%(100%)
E3 Rank 112 (6) 66%(66%) Rank 91 (42) 77%(77%)

weight, block 4

A4 Rank 273 (52) 66%(66%) Rank 161 (259) 44%(44%)
B4 Rank 2493 (457) 11%(77%) Rank 357 (203) 66%(44%)
C4 Rank 3501 (5163) 33%(33%) Rank 199 (899) 55%(55%)
D4 Rank 869 (886) 44%(55%) Rank 3247 (1965) 55%(66%)
E4 Rank 651 (580) 66%(11%) Rank 2684 (412) 55%(0%)


4.3.2 Discussion Experiment I

The comparison of the color distribution of all FPOIs for 2 colors with the color
distribution of the target pattern shows a direct connection between both distri-
butions - if white is the stronger color in the target pattern, the overall color dis-
tribution of all FPOIs for white is higher, and vice versa. For 3 colors it seemed
that subjects where attracted in all trials by mostly black areas. The second color
mostly depended on the stronger color in the target pattern color distribution. It
is to mention that all trials had a white background, and the used gray color had
a high contrast to this background. This means that dark colors had a higher
contrast compared to white. A high contrast is (following [9]) more appealing to
eye fixations, which is an explanation for this effect.
The analysis of the high similarity rankings for 2 colors showed very similar
patterns to the target pattern in the first placements of the list. For 3 colors the
results had no peculiar characteristics. The closer the FPOIs were to the target
pattern, the better were also the high similarity rankings.
Regarding the results for the sub-fixations, it seems that the subjects always
searched at least for a part of the pattern, in the case of 2 and 3 colors. Especially
for 3 colors with 81 possible sub-fixation patterns, the rankings showed that a
sub-fixation which was part of the target pattern was always on first or second
place.
Combined with the poor results in the high similarity rankings, this could
indicate that a pattern with 3 colors was to complex for a subject to look for.
Therefore, the subject searched only for a part of the target pattern. After this
part was found, the surrounding squares where compared with the whole target
pattern.
Regarding the high similarity rankings ordered by duration there are indica-
tors that the subject tended to look at patterns for a longer time which are similar
to the sub-pattern that is in his mind, but the rest of the observed pattern was
essentially different to the target pattern. It seems that subjects needed only a
short time period to recognize patterns which are really similar (e.g., match in
7 or 8 out of 9 positions). If only a part is similar, but the rest is more subtly
different, it takes longer for the subject to realize that there is a mismatch. E.g.
this is the case if the pattern on the screen is upside-down or rotated. Therefore,
the duration is an indicator for patterns that are somehow similar - either they
are rotated, or have parts in it which are similar to the part of the pattern the
subject is looking for - but it can not be said in which way this pattern is similar
(e.g. that the observed pattern is upside-down). Hence, no additional informa-
tion about the target pattern could be extracted out of the duration of a subject’s
fixation.
One big problem of experiment I was that the number of fixations and there-
fore the number of FPOIs was too low, so that its results could be impacted by


Figure 4.12: Patterns F1 and F2

Figure 4.13: Patterns G1 and G2

too unspecific eye movement data. (E.g., the first 30 patterns in a high similar-
ity ranking had all the same frequency, so the first 30 would have been all in
first place. ) Especially for 3 colors more eye movement data would have been
needed. Also, the subjects could scan the search picture in a systematical way to
find the target pattern which could influence the algorithm.
In order to avoid the systematical scanning and to receive more empirical
data, a second experiment with a smaller square size was conducted. Because
of the smaller square size the search picture was too big for subjects to scan it
systematically. Furthermore, more time and more eye movements were needed
to find the target pattern. With the second experiment the previous assumptions
should be validated.

4.3.3 Reports Experiment II
The main difference between experiment I and II is the smaller square size. The
square size is now 25x25 pixel, compared to 60x60 and 40x40 in experiment I.
For experiment II the number of fixations were much higher than in experiment
I.
Experiment II was run with 9 subjects. The two target patterns of block 1 and
2 for subject F are depicted in figure 4.12, for subject G in figure 4.13, for subject H
in figure 4.14, for subject I in figure 4.15, for subject J in figure 4.16, for subject K in
figure 4.17, for subject L in figure 4.18, for subject M in figure 4.19 and for subject
N in figure 4.20


Figure 4.14: Patterns H1 and H2

Figure 4.15: Patterns I1 and I2

Figure 4.16: Patterns J1 and J2

Figure 4.17: Patterns K1 and K2

Figure 4.18: Patterns L1 and L2

Figure 4.19: Patterns M1 and M2


Figure 4.20: Patterns N1 and N2


Color Distribution

In the discussion for experiment I (see section4.3.2) it was assumed that depend-
ing on the target pattern color distribution also the color distribution of the
FPOIs are influenced. Table 4.15 shows the color distribution for block 1 of ex-
periment II (2 colors), and table 4.16 the color distribution for block 2 (3 colors).

Table 4.15: Color distribution block 1 (Exp. 2)

F1 56%/44% 50%/50% 50%/50%
G1 67%/33% 51%/49% 51%/49%
H1 22%/78% 41%/59% 38%/62%
I1 44%/56% 49%/51% 49%/51%
J1 44%/56% 51%/49% 52%/48%
K1 89%/11% 57%/43% 59%/41%
L1 56%/44% 54%/46% 56%/44%
M1 33%/67% 48%/52% 48%/52%
N1 33%/67% 47%/53% 47%/53%

Table 4.16: Color distribution block 2 (Exp. 2)

F2 67%/22%/11% 38%/31%/31% 39%/30%/30%
G2 44%/22%/33% 36%/32%/31% 37%/32%/31%
H2 22%/44%/33% 30%/36%/33% 30%/37%/33%
I2 44%/44%/11% 32%/36%/31% 32%/37%/31%
J2 33%/44%/22% 31%/36%/33% 31%/36%/33%
K2 11%/44%/44% 30%/37%/32% 29%/38%/32%
L2 33%/11%/56% 40%/28%/31% 42%/26%/31%
M2 33%/33%/33% 34%/33%/32% 35%/32%/32%
N2 33%/56%/11% 30%/40%/30% 29%/42%/29%

For 2 colors, the color distribution of the target pattern does in fact influence
the color distribution of the FPOIs. The bigger the proportion of one color in the
pattern, the bigger it is in the color distribution.
For 3 colors the strongest color in the pattern was most of the times the
strongest color in the color distribution (except for M2).
Regarding M2: Looking at the top sub-fixations in the sub-fixation rankings the subject
probably looked for a part of the pattern which was mostly white. Therefore, the color
distribution of these FPOIs are dominated by white.


Fixation and High Similarity Rankings

Table 4.17 shows the high similarity ranking for 2 colors which most of the times
includes the target pattern in the top 3 ranks. Also, the average vector generated
of the first three places shows a high similarity to the target pattern (mostly 7 or
8 out of 9 accordance).


F1 Rank 3 100% Rank 12 66%
G1 Rank 1 100% Rank 2 100%
H1 Rank 3 77% Rank 2 77%
I1 Rank 2 88% Rank 2 88%
J1 Rank 2 88% Rank 1 88%
K1 Rank 1 88% Rank 2 88%
L1 Rank 2 77% Rank 2 77%
M1 Rank 13 77% Rank 18 77%
N1 Rank 1 88% Rank 1 77%

Table 4.18 disproves the same concept for 3 colors - the target pattern was
most of the time on a very low rank.


F2 Rank 180 88% Rank 871 33%
G2 Rank 6879 77% Rank 3584 33%
H2 Rank 837 44% Rank 3010 33%
I2 Rank 952 33% Rank 5079 33%
J2 Rank 9 55% Rank 37 33%
K2 Rank 1052 44% Rank 2015 44%
L2 Rank 694 55% Rank 497 44%
M2 Rank 7 66% Rank 205 33%
N2 Rank 999 55% Rank 711 55%

But besides the poor results for 3 colors, in some cases the high similarity
rankings show that the target pattern shape could be revealed - see figure 4.21
and figure 4.22. It seems that the high similarity ranking and its algorithm works
well for 2 colors - but with an additional color the algorithm does not show the
desired behavior. One reason could be the number of possible patterns for 3
colors - there are 19683 possible patterns, compared to 512 for 2 colors.


Figure 4.21: Pattern G2 fixation and high similarity ranking, radius 1

Sub-fixations

Table 4.19 shows that in trial block 1 the sub-fixation on first place in the sub-
fixation ranking was most of the time also part of the target pattern. If not, rank
2 was it then. Table 4.20 shows a similar result for trial block 2 with 3 colors - in
most of the times the top sub-fixation on rank one or two was also part of the
target pattern.

Duration of Fixations

The high similarity rankings ordered by duration are shown in table 4.21 for 2
colors and in table 4.22 for 3 colors. The high similarity rankings have become
worse when ordered by duration instead by frequency.


Figure 4.22: Pattern M2 ﬁxation and high similarity ranking, radius 1


Pattern Block 1 R1 Block 1 R0
F1 No (Rank 2) No (Rank 2)
G1 Yes Yes
H1 Yes Yes
I1 Yes Yes
J1 Yes Yes
K1 Yes Yes
L1 No (Rank 2) Yes
M1 No (Rank 2) No (Rank 2)
N1 Yes Yes


Pattern Block 2 R1 Block 2 R0
F2 Yes Yes
G2 No (Rank 2) No (Rank 2)
H2 No (Rank 10) No (Rank 9)
I2 No (Rank 3) Yes
J2 Yes Yes
K2 Yes No (Rank 2)
L2 No (Rank 3) No (Rank 2)
M2 Yes Yes
N2 Yes Yes


weight, block 1

F1 Rank 213 (3) 66%(100%) Rank 177 (12) 66%(66%)
G1 Rank 101 (1) 77%(100%) Rank 44 (2) 33%(100%)
H1 Rank 29 (3) 33%(77%) Rank 36 (2) 66(77%)
I1 Rank 2 (2) 88(88%) Rank 27 (2) 55%(88%)
J1 Rank 175 (2) 11%(88%) Rank 66 (1) 22%(88%)
K1 Rank 91 (1) 66%(88%) Rank 15 (2) 66%(88%)
L1 Rank 1 (2) 88%(77%) Rank 14 (2) 55%(77%)
M1 Rank 3 (13) 88%(77%) Rank 12 (18) 44%(77%)
N2 Rank 111(1) 88%(88%) Rank 37 (1) 22(77%)

weight, block 2

F2 Rank 2391 (180) 55%(88%) Rank 910 (871) 55%(33%)
G2 Rank 6565 (6879) 66%(77%) Rank 3347 (3584) 33%(33%)
H2 Rank 1037 (837) 44%(44%) Rank 3383 (3010) 55%(33%)
I2 Rank 9656 (952) 44%(33%) Rank 5577 (5079) 33%(33%)
J2 Rank 575 (9) 55%(55%) Rank 247 (37) 33%(33%)
K2 Rank 5854 (1052) 33%(44%) Rank 6442 (2015) 33%(44%)
L2 Rank 643 (694) 33%(55%) Rank 2471 (497) 44%(44%)
M2 Rank 326 (7) 11%(66%) Rank 593 (205) 66%(33%)
N2 Rank 113 (999) 55%(55%) Rank 207 (711) 66%(55%)


4.3.4 Discussion Experiment II
For 2 colors, the color distribution of the target pattern does in fact influence the
color distribution of the FPOIs. The bigger the proportion of one color in the
target pattern, the bigger it is in the color distribution of all FPOIs. For 3 colors
the strongest color in the target pattern was most of the times the strongest color
in the color distribution of the FPOIs.
It seems that the color distribution of the FPOIs and therefore the subjects’
fixations are mainly, but not only influenced by the target pattern. As already
mentioned in section 4.3.2 especially for 3 colors a high contrast is (following [9])
more appealing to eye fixations, so that black and gray could not really be dif-
ferentiated by the subject.
The assumption that a subject looks at similar patterns when searching for
the target pattern can be verified for 2 colors - the high similarity rankings for 2
colors show that the algorithm to create these rankings works well and that its
produced patterns are similar to the target pattern. For 3 colors the algorithm
is working poorly. It seems that a target pattern with 3 colors is too complex
for the human eye to search for it in a holistic way - eye movements and the
fixations are becoming more ”uncertain”.
The assumption that the duration of a fixation does not provide any addi-
tional information about the similarity to the target pattern can be verified - if
the high similarity list is ranked by the subject’s duration, the rankings become
worse. There are no additional information in the duration of a subject’s fixation
about the similarity to the target pattern.
The ranking for the sub-fixations proofed the idea that subjects search pri-
marily for a part of the target pattern. In the rankings, a sub-fixation which is
part of the target pattern is almost always in first or second place. Therefore,
the subject first searches for the sub-pattern; after he found a sub-pattern, he
compares the surroundings squares with the rest of the target pattern.
Furthermore, reducing the square size avoided the effect that subject’s could
scan the search picture systematically. Also, more fixations and therefore more
empirical data were generated.

Chapter 5

Conclusion

In this work an introduction into the field of eye tracking was given. A very
common task in eye tracking is to examine search behavior of subjects in differ-
ent search tasks - a lot of work has already been done in this area. Especially it
was found out that if a basic shape were given to subjects and they had to find
this in a ”random noise stimuli” (where the noise matched the average spectrum
of natural images) the eye movements tend to fixate areas that were similar to
the stimuli. An experiment was introduced were a target pattern and a stim-
uli consisting of squares in either two or three colors with different square sizes
were presented to the subject. The subjects were to find this target pattern in the
search picture. The question was if it is possible to calculate the target pattern
just with the given eye movements, without any knowledge of the target pattern
itself. (See chapter 2)
The software developed and used to create and run the experiment and
to analyze the output data was introduced and described in detail. The three
software parts named Experiment Builder, Experiment Engine and Experiment
Analysis Tool are the content of chapter 3.
The terms ROI and FPOI were explained. The ROI (region of interest) is the
area of surrounding squares of a subject’s fixation in the search picture. The size
of the ROI depends on a radius parameter. The FPOI (fixation pattern of interest) is
a pattern extracted from the ROI with the size of the target pattern. Furthermore,
the algorithms and parameters to create different rankings with the Experiment
Analysis Tools were described in detail.
Chapter 4 conducts an experiment with 14 subjects. The experiment was split
up into two different experiment setups and different block of trials for each
setup. The first experiment was conducted with 5 and the second experiment
with 9 participants. The main difference between the two experiments was the
smaller square size and because of this a higher number of fixations in the sec-

75

76 CHAPTER 5. CONCLUSION

ond experiment. Both experiments contained search task with two and three
colors. The experiment results and reports created with the Experiment Analy-
sis tool were summarized and discussed.
The results compare the color distribution of the target pattern with the over-
all color distrubtion of all FPOIs. Regarding the results of the experiments it can
be said that the color distribution of the target pattern influences the fixations
of the subject. Also the stronger color in the target pattern is the stronger color
in the color distribution of all FPOIs. For 3 colors more factors (like the con-
trast of the colors to the background) are involved, but the results show that the
strongest color in the target pattern is likely to be the strongest color in the color
distribution of all FPOIs.
A list of patterns that are very similar to each FPOI (8 out of 9 positions
in the pattern have the same color as the FPOI) is built. This list was ordered
by the amount of occurrences (frequency) of each pattern. For the experiments
with 2 colors, the list contained more often the target pattern as other patterns.
For 3 colors, the basic shape could be revealed in some results, but no special
correlation between the first places in the ranking and the target pattern could
be observed.
Creating the same ranking of patterns with high similarity, but ordering it by
the overall duration of the subjects’ fixation instead by frequency, the rank of the
target pattern in this ranking got worse. The duration of a subject fixation did
not contain additional information for the algorithm. That means that patterns
which the subject fixated for a longer time are not more similar to the target
pattern than patterns which the subject fixated for a shorter time period.
A list of all sub-fixations (a list of all 2x2 patterns inside a 3x3 FPOI) was
created and ranked by their frequency. Surprisingly, the results showed that
the sub-fixations on first places were in almost all experiments part of the target
pattern. Therefore, the subject memorized during his search task only a part of
the target pattern and searched for this part in the search pictures. When she/he
found this part in the picture, he/she then compared the surrounding squares
with the target pattern, and continued or finished searching.
Reducing the square size of the target pattern and search picture led to ”bet-
ter” results - by reducing the square size, a systematical scanning of the stimuli
by a subject was avoided. Furthermore, more fixations and therefore more em-
pirical data could be gathered, which improved the results of the algorithm.
Regarding the results for 3 colors, and compared to the results with 2 colors,
it can be concluded that 3 colors are too complex for a subject to search in a
holistic way. This means that the subject is not able to search for the pattern ”in
a whole” - he/she only looks for a part of the pattern, and then has to compare
the surroundings with the target. For 2 colors the subject probably can look for
the target pattern ”in a whole”, because it is less complex, so he/she can fixate

77

easier areas which are more similar.
For 2 colors the algorithm does produce good results, and also a smaller
square size led to improvements. Therefore, it can be assumed that the algo-
rithm works well with simple patterns that are black and white (or consist of
two colors with a high contrast) and that the calculated pattern is very similar
(or the same) to the target pattern. Furthermore, a smaller square size is advan-
tageous for the algorithm.
An application for this algorithm could be in image databases and content-
based image retrieval. Supposing a subject is very sensitive to certain patterns
or colors, the next time when she/he searches an image database, the algorithm
could be used to list the images which match to his search purposes.
Another application for the algorithm could be in the field of human-computer
interaction: there it is important to know the user’s intention. It could help the
user when he/she has problems finding a specific button or menu item. The
algorithm could help to generate suggestions that comply with his search pur-
pose.

Bibliography

[1] B.L. Beard and A.J. Ahumada. A technique to extract relevant image features for
visual tasks. Human Vision and Electronic Imaging III. SPIE Proceedings, 3299:79–85,
1998. Bellingham, WA: SPIE (In B. E. Rogowitz & T. N. Pappas). [cited at p. 10]

[2] D.C. Burr, M.C. Morrone, and J. Ross. Selective suppression of the magnocellu-
lar visual pathway during saccadic eye movements. Nature, 371:511Y513, 1994.
[cited at p. 6]

[3] SDL Community. Sdl (simple direct medialayer) website. http://www.libsdl.org.
[cited at p. 28]

[4] B. Efron. An introduction to the bootstrap. Boca Raton, FL: Chapman & Hall/CRC.,
1994. [cited at p. 11]

[5] J.M. Findlay. Saccade target selection during visual search. vision research. Vision
Research, 37:617Y631, 1997. [cited at p. 8]

[6] E. Hess and J. Polt. Pupil size as related to interest value of visual stimuli. Science,
140, 1960. [cited at p. 37]

[7] M.A. Just and P.A. Carpenter. The psychology of reading and language compre-
hension. Boston: Allyn & Bacon, 1987. [cited at p. 5]

[8] Universitt Konstanz. knime - konstanz information miner.
http://www.knime.org. [cited at p. 42]

[9] L.Itti and C. Koch. A saliency-based search mechanism for overt and covert shifts
of visual attention. Vision Research, 40(10-12):1489–1506, 2000. [cited at p. 64, 73]

[10] Microsoft. Microsoft download. http://www.microsoft.com/downloads/.
[cited at p. 29, 42]

[11] D.J. Parkhurst and E. Niebur. Scene content selected by active vision. Spatial Vision,
16:125–154, 2003. [cited at p. 9]

[12] M. Pomplun. Analysis and Models of Eye Movements in Comparative Visual Search.
PhD thesis, University Bielefeld, 1996. [cited at p. 6]

[13] M. Pomplun. Saccadic selectivity in complex visual search displays. Vision Re-
search, 46:1886–1900, 2006. [cited at p. 9, 12]

79

80 BIBLIOGRAPHY

[14] U. Rajashekar, A. Bovik, and L. Cormack. Visual search in noise: Revealing the in-
fluence of structural cues by gaze-contingent classification image analysis. Journal
of Vision, 6:379–386, 2006. [cited at p. 9, 10, 11, 12, 150]

[15] P. Reinagel and A.M. Zador. Natural scene statistics at the centre of gaze. Network,
10:341–350, 1999. [cited at p. 9]

[16] SR Research. Eye tracker ii overview. http://www.sr-
research.com/mount main.php. [cited at p. 7]

[17] SR Research. EyeLink Programmers Guide, version 3.0 edition. [cited at p. 28]

[18] SR Research. Sr research homepage. http://www.srresearch.com. [cited at p. 28]

[19] SR Research. EyeLink II User Manual, version 2.11 edition, 1997-2005. [cited at p. 26]

[20] R.F. Murray R.F and (et Al). Saccadic and perceptual performance in visual search
tasks: Ii. letter discrimination. Journal of the Optical Society of America A, Optics,
Image Science and Vision, 20:1356–1370, 2003. [cited at p. 8]

[21] Wikipedia. Eye tracking. http://en.wikipedia.org/wiki/Eye tracking, 07 2007.
[cited at p. 6]

[22] L.G. Williams. The effects of target specification on objects fixated during visual
search. Acta Psychologica, 27:355Y360, 1967. [cited at p. 8]

[23] J.M. Wolfe and T.S. Horowitz. What attributes guide the deployment of visual
attention and how do they do it? Neuroscience, 5(495Y501), 2004. [cited at p. 8]

[24] A.L. Yarbus. Eye movements and vision. New York Plenum Press, 1967. [cited at p. 5]

Appendix A

Used software and CD contents

A.1 Used Software
The following tools were used to develop and implement the code:

• Visual Studio 2005 Service Pack 2 from Microsoft
An IDE for developing C++ & .NET-Application. It includes Microsoft
.NET 2.0.

• Microsoft Office Primary Interop Assemblies (PIA) for Office
API for Microsoft Office, including Excel.

• Visual Paradigm for UML 6.1 Community Edition
An UML-Tool that integrates into Visual Studio and supports creating
UML diagrams.

• nDoc Beta 2005
An open source tool to create Documentation files from XML-documented
.NET-Sourcecode.

• Ghostdoc
A freeware Visual Studio plug-in to generate automated code documenta-
tion.

• kNime
Konstanz Information Miner, a data-mining tool.

• SR Research EyeTracker II Software Library
Tools and templates useful for developing EyeLink II experiments.

This work was created with Latex, using MikTex 2.6 and LEd (LatexEditor)
Version 0.51. Snapshots were made with MWSwap 3.

83

84 –

A.2 Contents of the Enclosed CD
• ASCLibrary
This folder contains the ASCLibrary project files and source codes, as de-
scribed as part of the Experiment Analysis Tool in section 3.3.3.

• BMPAnalysis
This folder contains the Experiment Analysis Tool project files and sources
codes, as described in section 3.3.3.

• BMPBuilder
This folder contains the Experiment Builder project files and source codes,
as described in section 3.1.3. Furthermore, the solution file that contains
the Experiment Analysis Tool, Experiment Builder and Experiment Engine
projects is inside this folder.

• BMPExperiment
This folder contains the Experiment Engine project files and sources codes,
as described in section 3.2.2.

• BMPExperimentData
The experiment created with the Experiment Builder, the eye movement
data from the EyeTracker device generated during the experiment runs as
well as the corresponding reports generated by the Experiment Analysis
Tool are (separated for each subject) inside this folder.

• Docs
The code documentation.

Appendix B

Namespace ASCLibrary

Namespace Contents Page

Classes
ASCBlink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Contains all information about a blink event in an ASC-Data File.
ASCData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Contains all information of events in an ASC-Data File, that is extracted
from an .edf-File of an EyeTracker experiment.
ASCFixation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Contains all information about a ﬁxation event in an ASC-Data File.
ASCSaccation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Contains all information about a saccade event in an ASC-Data File.
ASCTrialData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Contains the event information of a trial, namely all saccades, ﬁxations,
blinks and the trial metadata.
ASCTrialMetaData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Contains all metadata information of an ASC-File

85

86 ASCLibrary– ASCBlink

B.1 Classes
B.1.1 Class ASCBlink

Contains all information about a blink event in an ASC-Data File.

Declaration

public class ASCBlink
: Object

Properties

• Duration

public int Duration { get; set; }
Gets or sets the duration of the blink in ms.

• EndTime

public int EndTime { get; set; }
Gets or sets the end time of the blink in ms, relative to the start of the
experiment.

• StartTime

public int StartTime { get; set; }
Gets or sets the start time of the blink in ms, relative to the start of the
experiment.

Constructors

• .ctor

public ASCBlink( )

ASCLibrary– ASCData 87

B.1.2 Class ASCData

Contains all information of events in an ASC-Data File, that is extracted from an
.edf-File of an EyeTracker experiment.

Declaration

public class ASCData
: Object

Properties

• TrialDataList

public System.Collections.Generic.List{ASCLibrary.ASCTrialData}
TrialDataList { get; set; }
Gets or sets the trial data list.

Constructors

• .ctor

public ASCData( )

Methods

• LoadData

public void LoadData( )
Loads the data of an ASC-File.

– Parameters
∗ filename - The ﬁlename.

88 ASCLibrary– ASCFixation

B.1.3 Class ASCFixation

Contains all information about a fixation event in an ASC-Data File.

Declaration

public class ASCFixation
: Object

Properties

• AveragePupilSize

public int AveragePupilSize { get; set; }
Gets or sets the average size of the pupil in pixel.

• AverageX

public int AverageX { get; set; }
Gets or sets the average X fixation value on the screen.

• AverageY

public int AverageY { get; set; }
Gets or sets the average Y fixation value on the screen.

• Duration

Gets or sets the duration of the fixation in ms.

• EndTime

Gets or sets the end time of the fixation in ms, relative to the start of the
experiment.

ASCLibrary– ASCFixation 89

• PosX

public int PosX { get; set; }
Gets or sets the X-Position of the fixation. Not used/deprecated.

• PosY

public int PosY { get; set; }
Gets or sets the Y-Pos of the fixation. Not used/deprecated

• StartTime

Gets or sets the start time of the fixation in ms, relative to the start of the
experiment.

Constructors

• .ctor

public ASCFixation( )

90 ASCLibrary– ASCSaccation

B.1.4 Class ASCSaccation

Contains all information about a saccade event in an ASC-Data File.

Declaration

public class ASCSaccation
: Object

Properties

• Duration

Gets or sets the duration in ms.

• EndTime

Gets or sets the end time in ms, relative to the start of the experiment.

• EndX

public int EndX { get; set; }
Gets or sets the end X-position on the screen.

• EndY

public int EndY { get; set; }
Gets or sets the end Y-position on the screen.

• PeakVelocity

public int PeakVelocity { get; set; }
Gets or sets the peak velocity.

ASCLibrary– ASCSaccation 91

• SaccadicAmplitude

public decimal SaccadicAmplitude { get; set; }
Gets or sets the saccadic amplitude.

• StartTime

Gets or sets the start time in ms, relative to the start of the experiment.

• StartX

public int StartX { get; set; }
Gets or sets the start X-position on the screen.

• StartY

public int StartY { get; set; }
Gets or sets the start Y-position on the screen.

Constructors

• .ctor

public ASCSaccation( )

92 ASCLibrary– ASCTrialData

B.1.5 Class ASCTrialData

Contains the event information of a trial, namely all saccades, ﬁxations, blinks and
the trial metadata.

Declaration

public class ASCTrialData
: Object

Properties

• BlinkList

public System.Collections.Generic.List{ASCLibrary.ASCBlink} BlinkList
{ get; set; }

Gets or sets the blink list.

• FixationList

public System.Collections.Generic.List{ASCLibrary.ASCFixation}
FixationList { get; set; }

Gets or sets the ﬁxation list.

• SaccationList

public System.Collections.Generic.List{ASCLibrary.ASCSaccation}
SaccationList { get; set; }

Gets or sets the saccation list.

• TrialMetaData

public ASCLibrary.ASCTrialMetaData TrialMetaData { get; set; }

Gets or sets the trial meta data.

ASCLibrary– ASCTrialData 93

Constructors

• .ctor

public ASCTrialData( )

94 ASCLibrary– ASCTrialMetaData

B.1.6 Class ASCTrialMetaData

Contains all metadata information of an ASC-File

Declaration

public class ASCTrialMetaData
: Object

Properties

• EndTime

Gets or sets the end time.

• StartTime

Gets or sets the start time.

• Timeout

public bool Timeout { get; set; }
Gets or sets a value indicating whether this trial timed out.

• TimeoutTime

public int TimeoutTime { get; set; }
Gets or sets the time after which the time out happened.

• TrialID

public int TrialID { get; set; }
Gets or sets the trial ID.

BMPAnalysis– ASCTrialMetaData 95

• TrialName

public string TrialName { get; set; }
Gets or sets the name of the trial.

Constructors

• .ctor

public ASCTrialMetaData( )

Appendix C

Namespace BMPAnalysis


Classes
AnalysisObject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
This object contains all data needed to generate an analysis.
ColorDistribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Defines a color distribution of two or more colors - a distribution contains
the percentage.
Config . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Represents a configuration file which contains the resolution of the display
screen and the overall number of trials of an experiment
Info . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Displays the Info window with description about the author and program.

MainForm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Displays the main GUI for the analysis tool.
OutputFile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Analyzes the data and makes all the necessary calculations to create the
analysis output file - to implement this class, the abstract methods should
generate the equivalent output for its specified file format.
OutputFileExcel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Creates and displays excel sheets with the data output.
OutputFileParameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Represents the parameter for the calculation of the analysis and creating the
output files.
OutputFileRawData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

97

98 BMPAnalysis– ASCTrialMetaData

Creates and display text files that contain raw data with the fixations and
results of the analysis.
Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
A vector is a one-dimensional representation of a two dimensional pattern.
If the pattern is e.g. a 3x3 array of different color values, a vector of this pat-
tern would be an one-dimensional array with the length 9. The first three
color values of the vector array would be identical with the top three color
values of the pattern array.(The top left, top center and top right colors).
The second group of three would correspond to the colors in the middle
row, and the the last three values would be the ones in the bottom row.
Therefore, the conversion is row wise from top to bottom. A vector can be
written as a string by writing the abbreviated color letters in a single row
(E.g.: ”WWWB” is a vector string for describing a vector with the color
sequence white,white,white and black.)

BMPAnalysis– AnalysisObject 99

C.1 Classes
C.1.1 Class AnalysisObject

This object contains all data needed to generate an analysis.

Declaration

public class AnalysisObject
: Object

Properties

• AscData

public ASCLibrary.ASCData AscData { get; set; }
Gets or sets the asc data.

• Config

public BMPAnalysis.Config Config { get; set; }
Gets or sets the config data.

• DataFile

public string DataFile { get; set; }
Gets or sets the data file.

• Experiment

public BMPBuilder.ExperimentData Experiment { get; set; }
Gets or sets the experiment data.

• ExperimentFolder

public string ExperimentFolder { get; set; }
Gets or sets the experiment folder.

100 BMPAnalysis– AnalysisObject

• Name

public string Name { get; set; }
Gets or sets the name.

• ParameterList

public BMPAnalysis.OutputFileParameter ParameterList { get; set; }
Gets or sets the parameter list.

• PatternList

public System.Collections.Generic.List{BMPBuilder.Pattern} PatternList
{ get; set; }
Gets or sets the pattern list.

• PictureList

public System.Collections.Generic.List{BMPBuilder.Picture} PictureList {
get; set; }
Gets or sets the picture list.

Constructors

• .ctor

public AnalysisObject( )

BMPAnalysis– ColorDistribution 101

C.1.2 Class ColorDistribution

Deﬁnes a color distribution of two or more colors - a distribution contains the per-
centage.

Declaration

public class ColorDistribution
: Object

Properties

• Frequency

public long Frequency { get; set; }
Gets or sets the frequency.

• Percentage

public double Percentage { get; set; }
Gets or sets the percentage.

Constructors

• .ctor

public ColorDistribution( )
Initializes a new instance of the class.

– Parameters
∗ percentage - The percentage.
∗ frequency - The frequency.

Methods

102 BMPAnalysis– ColorDistribution

• CompareTo

public int CompareTo( )
Compares the current instance with another object of the same type.

– Parameters
∗ obj - An object to compare with this instance.

BMPAnalysis– Config 103

C.1.3 Class Config

Represents a configuration file which contains the resolution of the display screen
and the overall number of trials of an experiment

Declaration

public class Config
: Object

Properties

• DisplayX

public int DisplayX { get; set; }
Gets or sets the width of the display.

• DisplayY

public int DisplayY { get; set; }
Gets or sets the height of the display.

• NumTrials

public int NumTrials { get; set; }
Gets or sets the overall number of trials.

Constructors

• .ctor

public Config( )

• .ctor

104 BMPAnalysis– Config

public Config( )

– Parameters
∗ filename - The filename of the config file to be load.

Methods

• Load

public void Load( )
Loads the specified config filef.

– Parameters
∗ filename - The filename of the config file.

BMPAnalysis– Info 105

C.1.4 Class Info


Declaration

public class Info
: Form

Constructors

• .ctor

public Info( )

106 BMPAnalysis– MainForm

C.1.5 Class MainForm

Displays the main GUI for the analysis tool.

Declaration

public class MainForm
: Form

Constructors

• .ctor

public MainForm( )

BMPAnalysis– OutputFile 107

C.1.6 Class OutputFile

Analyzes the data and makes all the necessary calculations to create the analysis
output file - to implement this class, the abstract methods should generate the equiva-
lent output for its specified file format.

Declaration

public class OutputFile
: Object

Properties

• AscData

public ASCLibrary.ASCData AscData { get; set; }
Gets or sets the ASC data.

• Config

public BMPAnalysis.Config Config { get; set; }
Gets or sets the config data.

• DataFile

public string DataFile { get; set; }
Gets or sets the data file.

• Experiment

Gets or sets the experiment.

• ExperimentFolder

public string ExperimentFolder { get; set; }
Gets or sets the experiment folder.

108 BMPAnalysis– OutputFile

• Parameter

public BMPAnalysis.OutputFileParameter Parameter { get; set; }
Gets or sets the parameter.

• PatternList

public System.Collections.Generic.List{BMPBuilder.Pattern} PatternList
{ get; set; }
Gets or sets the pattern list.

• PatternVector

public BMPAnalysis.Vector PatternVector { get; set; }
Gets or sets the pattern vector.

• PictureList

public System.Collections.Generic.List{BMPBuilder.Picture} PictureList {
get; set; }
Gets or sets the picture list.

Constructors

• .ctor

public OutputFile( )

– Parameters
∗ parameterList - The parameter list.
∗ experiment - The experiment data.
∗ pictureList - The picture list.
∗ patternList - The pattern list.
∗ config - The config data.
∗ ascData - The ASC data.
∗ dataFile - The data file.
∗ experimentFolder - The experiment folder.

BMPAnalysis– OutputFile 109

Methods

• CreateOutputFile

public void CreateOutputFile( )
Creates the output ﬁle.

– Parameters
∗ worker - The background worker object, to which progress is re-
ported.

110 BMPAnalysis– OutputFileExcel

C.1.7 Class OutputFileExcel

Creates and displays excel sheets with the data output.

Declaration

public class OutputFileExcel
: OutputFile

Constructors

• .ctor

public OutputFileExcel( )

– Parameters

BMPAnalysis– OutputFileParameter 111

C.1.8 Class OutputFileParameter

Represents the parameter for the calculation of the analysis and creating the output
files.

Declaration

public class OutputFileParameter
: Object

Properties

• Excel

public bool Excel { get; set; }
Gets or sets a value indicating whether the output should be an Excel
sheet.

– Usage
∗ Defines a XOR-Relationship between the excel and the rawData
Flag.

• From

public int From { get; set; }
Gets or sets at which trial to start the output.

• IncludeFirstFixation

public bool IncludeFirstFixation { get; set; }
Gets or sets a value indicating whether the first fixation of a trial should be
included.

• IncludePatternFixation

public bool IncludePatternFixation { get; set; }
Gets or sets a value indicating whether the pattern fixation should be in-
cluded or not.

112 BMPAnalysis– OutputFileParameter

• MoveX

public int MoveX { get; set; }
Gets or sets if and where the fixation position should be moved in vertical
direction. For instance, a value of -1 would move the fixation 1 square to
the left.

• MoveY

public int MoveY { get; set; }
Gets or sets if and where the fixation position should be moved in horizon-
tal direction. For instance, a value of -1 would move the fixation 1 square
up.

• OutputFolder

public string OutputFolder { get; set; }
Gets or sets the output folder. Only signification if raw data (txt) files are
desired.

• Radius

public int Radius { get; set; }
Gets or sets the radius. Indicates how many surrounding squares of a
fixation (additional to the pattern size) should be included to the fixation
calculation.

• RawData

public bool RawData { get; set; }
Gets or sets a value indicating whether the output should be raw data (txt)
files.

– Usage
∗ Defines a XOR-Relationship between the excel and the rawData
Flag.

• ShortExcelReport

BMPAnalysis– OutputFileParameter 113

public bool ShortExcelReport { get; set; }
Gets or sets a value indicating whether a short excel report should be gen-
erated.

• SimilarityMeasure

public double SimilarityMeasure { get; set; }
Gets or sets the similarity measure, which is a percentage value that de-
scribes how similar two vectors are to each other.

• To

public int To { get; set; }
Gets or sets at which trial to end the output.

• Top10

public int Top10 { get; set; }
Gets or sets how many vectors should be included in the ﬁnal ranking list.

Constructors

• .ctor

public OutputFileParameter( )

Methods

• ToString

public string ToString( )
Returns a String that represents the current OutputFileParameter.

114 BMPAnalysis– OutputFileRawData

C.1.9 Class OutputFileRawData

Creates and display text ﬁles that contain raw data with the ﬁxations and results of
the analysis.

Declaration

public class OutputFileRawData
: OutputFile

Constructors

• .ctor

public OutputFileRawData( )

– Parameters

BMPAnalysis– Vector 115

C.1.10 Class Vector

A vector is a one-dimensional representation of a two dimensional pattern. If the
pattern is e.g. a 3x3 array of different color values, a vector of this pattern would be an
one-dimensional array with the length 9. The ﬁrst three color values of the vector array
would be identical with the top three color values of the pattern array.(The top left, top
center and top right colors). The second group of three would correspond to the colors
in the middle row, and the the last three values would be the ones in the bottom row.
Therefore, the conversion is row wise from top to bottom. A vector can be written as a
string by writing the abbreviated color letters in a single row (E.g.: ”WWWB” is a vector
string for describing a vector with the color sequence white,white,white and black.)

Declaration

public class Vector
: Object

Properties

• ColorList

public System.Collections.Generic.List{System.Drawing.Color} ColorList
{ get; set; }
Gets or sets a list of all colors the vector can (but may not) contain.

• Colors

public System.Int32[] Colors { get; }
Gets the color values as integers of the vector array.

• Frequency

public long Frequency { get; set; }
Gets or sets the frequency. If a number of vectors are combined in a list,
the frequency describes the occurrences in this list.

• Size

116 BMPAnalysis– Vector

public int Size { get; set; }
Gets or sets the size of the vector. This is basically the length of the vector
array.

• Weight

public long Weight { get; set; }
Gets or sets the weight of the vector. This can be the duration of a ﬁxation.

Constructors

• .ctor

public Vector( )

– Parameters
∗ size - The size of the vector. This is basically the length of the
vector array.
∗ colorList - A list of all colors the vector can (but may not) contain.

• .ctor

public Vector( )

– Parameters
∗ colorArray -
∗ colorList -

• .ctor

public Vector( )

– Parameters
∗ size - The size of the vector. This is basically the length of the
vector array.
∗ weight - The weight of the vector, e.g. the duration of a ﬁxation.


∗ colorList - A list of all colors the vector can (but may not) contain.

• .ctor

public Vector( )

– Parameters
∗ vector - An existing vector that values will be copied into the new
vector object, except its frequency and its weight.
∗ weight - The new weight of the new vector.
∗ frequency - The new frequency of the new vector.

Methods

• Add

public void Add( )
Adds the specified color to the vector. If the vector is newly instantiated,
this will start at the first element of the vector, and continues with the next
element, until the vector is filled.

– Parameters
∗ color - The color value as integer that should be added at the next
unset vector position.

• Add

public void Add( )
Adds the specified color to the vector. If the vector is newly instantiated,
this will start at the first element of the vector, and continues with the next
element, until the vector is filled.

– Parameters
∗ color - The color value as color object that should be added at the
next unset vector position.


• AverageVector

public BMPAnalysis.Vector AverageVector( )
Calculates a new average vector of a list of vectors. Counts the frequency
(or the total weight) of each color for each position in all vectors, and writes
the color with the highest frequency (or highest weight) in the new average
vector. If the frequency for a position is the same for all colors, the color
which has color the highest value in the color distribution is taken.

– Parameters
∗ vectorList - The vector list to build the new average vector of.
∗ range - The range, how many of the vectors in the vector list
should be considered. If the value is smaller than the size of the
vector list, the ﬁrst vectors up to the given range of the list will
be taken
∗ colorList - The list of all possible colors as an integer array that
can, but may not occur in the vectors.
∗ colorDistribution - The color distribution of all possible colors
that can, but may not occur in the vectors.
∗ useWeight - if set to true, the weight instead the frequency is used
to build the average vector.

• AverageVector


– Parameters
∗ colorList - The list of all possible colors as a list of color objects


• AverageVector


– Parameters
∗ colorList - The list of all possible colors as an integer array that
can, but may not occur in the vectors.

• AverageVector


– Parameters
∗ range - The range, how many of the vectors in the vector list
should be considered. If the value is smaller than the size of the
vector list, the ﬁrst vectors up to the given range of the list will
be taken
∗ colorList - The list of all possible colors as a list of color objects


• ColorDistribution

public System.Double[] ColorDistribution( )
Calculates the percentage of the color distribution.

– Parameters
∗ decimals - How many decimal places the percentage should
have.

• ColorDistributionString

public string ColorDistributionString( )
Returns a string which describes the percentage of the color distrubution.

– Parameters
∗ decimals - How many decimal places the percentage should
have.

• Compare

public double Compare( )
Calculates the similarity of the current instance with another vector. The
similarity is a percentage which indicates how many positions have the
same value in both vectors.

– Parameters
∗ v1 - The vector to compare the current instance with.

• Compare

public double Compare( )
Calculates the similarity of two vectors with each other. The similarity is
a percentage which indicates how many positions have the same value in
both vectors.

– Parameters
∗ v1 - The ﬁrst vector to compare with the second vector.
∗ v2 -


• CompareTo

public int CompareTo( )
Compares the current instance with another vector object. Uses the fre-
quency to compare these vectors.

– Parameters
∗ obj - An object to compare with this instance.

• CompareToUsingWeight

public int CompareToUsingWeight( )
Compares two vectors with each other. Uses the weight to compare these
vectors.

– Parameters
∗ obj1 - The ﬁrst vector
∗ obj2 - The second vector.

• CreatePossibleVectors

public System.Collections.Generic.List{BMPAnalysis.Vector}
CreatePossibleVectors( )
Creates a list of all possible vectors that can be created with the given
amount of colors.

– Parameters
∗ colors - The possible colors that can, but may not appear in each
vector.
∗ dimX - The width of the pattern. The size of the vector will be the
product of the given width and height of the pattern.
∗ dimY - The height of the pattern. The size of the vector will be
the product of the given width and height of the pattern.

• Equals

public bool Equals( )
Determines whether the speciﬁed Vector is equal to the current Vector.

– Parameters

122 BMPBuilder– Vector

∗ obj - The Vector to compare with the current Vector.

• Swap

public bool Swap( )
Swaps the colors of two positions in the vector array.

– Parameters
∗ pos1 - The ﬁrst position in the vector array to be swapped with
the second position.
∗ pos2 -

• ToString

public string ToString( )
Returns a (a vector string) that represents the current .

Appendix D

Namespace BMPBuilder


Classes
Bitmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
The Bitmap class defines basic properties and functions for a square-shaped
checkboard-like image. Creating, Painting, Loading and Saving of images
is possible. This class should not be instantiated directly - its inherited
classes Picture and Pattern should be used instead.
Constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
The Constants class defines fixed values for the experiments, such as the
screen size or the maximum number of squares for a pattern.
ExperimentData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Class to store and handle all necessary information for an experiment. It
contains a list of all Trials as well as functions to load and save the experi-
ment and generate data files for the experiment engine.
ExperimentForm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
The GUI for the Experiment Builder.
Info . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
The Pattern class inherits from the Bitmap class, and represents a pattern
image.
Picture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
The Picture class inherits from the Bitmap class, and represents a picture in
which a target pattern can be found.
PreviewData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

123

124 BMPBuilder– Vector

Class to store and handle all necessary information of a trial which is
needed to generate a preview of a trial.
PreviewForm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
This class displays a rough preview of a pattern and picture on the screen.
TrialData . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Class to store and handle all necessary information of a trial. All trial data
which is need to generate a preview is stored the linked PreviewData object.

TrialForm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .147
This class is the GUI for entering trial information.

BMPBuilder– Bitmap 125

D.1 Classes
D.1.1 Class Bitmap

The Bitmap class deﬁnes basic properties and functions for a square-shaped
checkboard-like image. Creating, Painting, Loading and Saving of images is possible.
This class should not be instantiated directly - its inherited classes Picture and Pattern
should be used instead.

Declaration

public class Bitmap
: Object

Properties

• Colors

public System.Collections.Generic.List{System.Drawing.Color} Colors {
get; set; }
A list of all colors the Image contains. (E.g., if the Image contains squares
in black and white, this List will contain two color objects with the values
black and white.)

• SquaresX

public int SquaresX { get; set; }
Numbers of squares in horizontal directions

• SquaresY

public int SquaresY { get; set; }
Numbers of sqares in vertical direction

• StartLocation

public System.Drawing.Point StartLocation { get; set; }

126 BMPBuilder– Bitmap

The X and Y coordinate the image is displayed on the screen. Important to
calcuate a square Position from a given point on the screen.

• Structure

public System.Int32[,] Structure { get; set; }
The structure of the image, deﬁned in a two-dimensional array of integer
values.

Constructors

• .ctor

public Bitmap( )
Creates a new image object.

– Parameters
∗ squaresX - The number of squares in horizontal direction.
∗ squaresY - The number of squares in vertical direction.
∗ colors - A list of colors which should be used for the image. (E.g,
black and white)

Methods


Returns the percentage distribution of colors of an (already created) im-
age. E.g, if a 2x2 image consists of 1 black square and 3 white squares, the
distribution would be 0.25 and 0.75.

– Parameters
∗ decimals - Truncates the result to the speciﬁed decimal places



BMPBuilder– Bitmap 127

Returns the percentage distribution of colors of the defined clipping area
of an image. E.g, if the defined clipping area spans over a 2x2 pattern and
consists of 1 black square and 3 white squares, the distribution would be
0.25 and 0.75.

– Parameters
∗ decimals - Truncates the result to the specified decimal places
∗ distRect - A clipped area defined as a rectangle. Should not ex-
ceed the size of the image.

• Create

public void Create( )
Fills the image at random with the colors defined in the Color Property

• Load

public BMPBuilder.Bitmap Load( )
Loads previously saved image data.

– Parameters
∗ filename - The path and filename of a previously saved object.

• Paint

public System.Drawing.Bitmap Paint( )
Returns a ”real” bitmap of the type System.Drawing.Bitmap of the created
image.

– Parameters
∗ squareSize - The Width and Height of each Square in Pixel
∗ borderColor - The color of the surrounding border

• Save

public void Save( )
Saves the image data to a file on the hardisk.

– Parameters
∗ filename - The path and filename where to store the object.

128 BMPBuilder– Constants

D.1.2 Class Constants

The Constants class defines fixed values for the experiments, such as the screen size
or the maximum number of squares for a pattern.

Declaration

public class Constants
: Object

Properties

• BorderSize

public int BorderSize { get; set; }
Gets or sets the size of the rectangle frame that marks the target squares at
the end of each trial.

• ConfigFileName

public string ConfigFileName { get; }
Gets the name of the config file.

• DataFolderName

public string DataFolderName { get; }
Gets the name of the data folder.

• DataPostfix

public string DataPostfix { get; }
Gets the data file ending (e.g.,set to .bin)

• DisplayX

public int DisplayX { get; }
Gets the horizontal resolution of the experiment screen.

BMPBuilder– Constants 129

• DisplayY

public int DisplayY { get; }
Gets the vertical resolution of the experiment screen.

• ExperimentFileName

public string ExperimentFileName { get; }
Gets the name of the experiment file.

• ImageFolderName

public string ImageFolderName { get; }
Gets the name of the image folder.

• MaxPatternsX

public int MaxPatternsX { get; }
Gets the maximum number of pattern squares in horizontal direction.

• MaxPatternsY

public int MaxPatternsY { get; }
Gets the maximum number of pattern squares in vertical direction.

• MaxSizeX

public int MaxSizeX { get; }
Gets the maximum width for a square.

• MaxSizeY

public int MaxSizeY { get; }
Gets the maximum height for a square.

• PatternDataPrefix

public string PatternDataPrefix { get; }


Gets the pattern data prefix.

• PatternHeight

public int PatternHeight { get; }

• PatternNumberBeforeTimeChange

public int PatternNumberBeforeTimeChange { get; }
Gets the number of trials done before display time changes for first time.

• PatternNumberBeforeTimeChange2

public int PatternNumberBeforeTimeChange2 { get; }
Gets the number of trials done before display time changes for second
time.

• PatternOffset

public int PatternOffset { get; }
Gets the vertical (X-Axis) start position of the pattern display.

• PatternWidth

public int PatternWidth { get; }
Gets the maximum width for a pattern object.

• PictureDataPrefix

public string PictureDataPrefix { get; }
Gets the picture data prefix.

• PictureHeight

public int PictureHeight { get; }
Gets the maximum height for a picture object.

• PictureWidth

BMPBuilder– Constants 131

public int PictureWidth { get; }
Gets the maximum width for a picture object.

Constructors

• .ctor

public Constants( )

Methods

• CalcPatternLocation

public System.Drawing.Point CalcPatternLocation( )
Calculates the pattern location on the screen.

– Parameters
∗ sameScreen - If set to true, pattern and picture are displayed on
the same screen.
∗ patternSquaresX -
∗ sizeX - The size of each square (in pixel) in horizontal direction.
∗ patternSquaresY -
∗ sizeY - The size of each square (in pixel) in vertical direction.

• CalcPictureLocation

public System.Drawing.Point CalcPictureLocation( )
Calculates the picture location on the screen.

– Parameters
∗ sameScreen - If set to true, pattern and picture are displayed on
the same screen.
∗ sizeX - The size of each square (in pixel) in horizontal direction.
∗ sizeY - The size of each square (in pixel) in vertical direction.


• ColorIntToString

public string ColorIntToString( )
Makes a string out of a color object.

– Parameters
∗ colValue - The color value as an object.
∗ abbr - If set to true, an abbreviated form will be used.

• ColorIntToString

public string ColorIntToString( )
Makes a string out of an integer color value.

– Parameters
∗ colValue - The color value as an integer.
∗ abbr - If set to true, an abbreviated form will be used.

BMPBuilder– ExperimentData 133

D.1.3 Class ExperimentData

Class to store and handle all necessary information for an experiment. It contains a
list of all Trials as well as functions to load and save the experiment and generate data
files for the experiment engine.

Declaration

public class ExperimentData
: Object

Properties

• Trials

public System.Collections.Generic.List{BMPBuilder.TrialData} Trials {
get; set; }
Gets or sets the list of trials.

Constructors

• .ctor

public ExperimentData( )

Methods

• GenerateExperimentFile

public bool GenerateExperimentFile( )
Generates the experiment file for the experiment engine..

– Parameters
∗ folder - The folder where the files are to be created.
∗ worker - A background worker object that runs this function in a
background thread and can retrieve progress information.

134 BMPBuilder– ExperimentData

• LoadExperiment

public BMPBuilder.ExperimentData LoadExperiment( )
Loads the experiment.

– Parameters
∗ filename - The ﬁlename (including path) of the experiment to be
loaded.

• SaveExperiment

public void SaveExperiment( )
Saves the experiment.

– Parameters
∗ filename - The ﬁlename (including path) of the experiment.

BMPBuilder– ExperimentForm 135

D.1.4 Class ExperimentForm

The GUI for the Experiment Builder.

Declaration

public class ExperimentForm
: Form

Properties

• Experiment

Gets or sets the experiment data.

Constructors

• .ctor

public ExperimentForm( )

Methods

• UpdateTrialList

public void UpdateTrialList( )
Updates the trial list.

136 BMPBuilder– Info

D.1.5 Class Info


Declaration

public class Info
: Form

Constructors

• .ctor

public Info( )

BMPBuilder– Pattern 137

D.1.6 Class Pattern

The Pattern class inherits from the Bitmap class, and represents a pattern image.

Declaration

public class Pattern
: Bitmap

Constructors

• .ctor

public Pattern( )

– Parameters
black and white)

138 BMPBuilder– Picture

D.1.7 Class Picture

The Picture class inherits from the Bitmap class, and represents a picture in which a
target pattern can be found.

Declaration

public class Picture
: Bitmap

Properties

• PatternPositionX

public int PatternPositionX { get; set; }
Gets or sets the pattern position X-coordinate (horizontal).

• PatternPositionY

public int PatternPositionY { get; set; }
Gets or sets the pattern position Y-coordinate (vertical).

Constructors

• .ctor

public Picture( )

– Parameters
black and white)

BMPBuilder– Picture 139

Methods

• CalculatePosX

public int CalculatePosX( )
Calculates the horizontal position in the structure of a given pixel location
on the screen. Depends on the start location and the size of the squares.

– Parameters
∗ averageX - The pixel location (X-Axis/horizontal) on the screen.
∗ sizeX - The horizontal size of each square.

• CalculatePosY

public int CalculatePosY( )
Calculates the vertical position in the structure of a given pixel location on
the screen. Depends on the start location and the size of the squares.

– Parameters
∗ averageY -
∗ sizeY -

• IsPatternFixation

public bool IsPatternFixation( )
Determines whether the given position in the structure is a pattern fixa-
tion. Includes the surrounding squares - the radius specifies how many
squares are included.

– Parameters
∗ pos - The position in the structure.
∗ radius - The radius, specifies how many surrounding squares
are included. A radius of 0 includes only the minimun squares,
which depends on the pattern size. The bigger the radius, the
more squares (also depending on the pattern size) are included.
∗ includePatternFixation - If set to true the function always returns
false.

140 BMPBuilder– Picture

• IsPatternFixation

public bool IsPatternFixation( )
Determines whether the given position in the structure is a pattern fixa-
tion. Includes the surrounding squares - the radius specifies how many
squares are included.

– Parameters
∗ pos - The position in the structure.
∗ radius - The radius, specifies how many surrounding squares
are included. A radius of 0 includes only the minimun squares,
which depends on the pattern size. The bigger the radius, the
more squares (also depending on the pattern size) are included.

• RemoveAndAddPattern

public void RemoveAndAddPattern( )
Removes and adds a pattern to the picture.

– Parameters
∗ pattern - The pattern to be added.

BMPBuilder– PreviewData 141

D.1.8 Class PreviewData

Class to store and handle all necessary information of a trial which is needed to
generate a preview of a trial.

Declaration

public class PreviewData
: Object

Properties

• BackgroundColor

public System.Drawing.Color BackgroundColor { get; set; }
Gets or sets the color of the background.

• BorderColor

public System.Drawing.Color BorderColor { get; set; }
Gets or sets the color of the border.

• Color1

public System.Drawing.Color Color1 { get; set; }
Gets or sets the color 1.

• Color2


• Color3


142 BMPBuilder– PreviewData

• Color3Checked

public bool Color3Checked { get; set; }
Gets or sets a value indicating whether color 3 is activated.

• Color4


• Color4Checked


• Color5


• Color5Checked


• Color6


• Color6Checked


• ConclusionColor

public System.Drawing.Color ConclusionColor { get; set; }

BMPBuilder– PreviewData 143

Gets or sets the color of the rectangular frame the marks the target pattern
in the picture at the end of a trial.

• DisplayX

public int DisplayX { get; set; }
Gets or sets the width of the display - should be the screen width (in pixel)
of the experiment pc.

• DisplayY

public int DisplayY { get; set; }
Gets or sets the height of the display - should be the screen width (in pixel)
of the experiment pc.

• PatternDisplayTime

public int PatternDisplayTime { get; set; }
Gets or sets the pattern display time for the first trials.

• PatternDisplayTime2

public int PatternDisplayTime2 { get; set; }
Gets or sets the pattern display time after a specified number of trials.

• PatternDisplayTime3

public int PatternDisplayTime3 { get; set; }
Gets or sets the pattern display time after a specified number of trials.

• PatternSquaresX

public int PatternSquaresX { get; set; }
Gets or sets the number of squares in horizontal direction for the pattern.

• PatternSquaresY

public int PatternSquaresY { get; set; }
Gets or sets the number of squares in vertical direction for the pattern.

144 BMPBuilder– PreviewData

• SameScreen

public bool SameScreen { get; set; }
Gets or sets a value indicating whether the pattern and the picture are
displayed on the same screen.

• SizeX

public int SizeX { get; set; }
Gets or sets the width of each square (in pixel).

• SizeY

public int SizeY { get; set; }
Gets or sets the height of each square (in pixel).

• SquaresX

public int SquaresX { get; set; }
Gets or sets the number of squares in horizontal direction in a trial.

• SquaresY

public int SquaresY { get; set; }
Gets or sets the number of squares in vertical direction in a trial.

Constructors

• .ctor

public PreviewData( )

BMPBuilder– PreviewForm 145

D.1.9 Class PreviewForm

This class displays a rough preview of a pattern and picture on the screen.

Declaration

public class PreviewForm
: Form

Constructors

• .ctor

public PreviewForm( )

Methods

• CreatePreviewForm

public void CreatePreviewForm( )
Creates the preview form.

– Parameters
∗ preview - The preview data.

146 BMPBuilder– TrialData

D.1.10 Class TrialData

Class to store and handle all necessary information of a trial. All trial data which is
need to generate a preview is stored the linked PreviewData object.

Declaration

public class TrialData
: Object

Properties

• Duration

Gets or sets the duration of the trial.

• Name

public string Name { get; set; }
Gets or sets the name of the trial.

• Preview

public BMPBuilder.PreviewData Preview { get; set; }
Gets or sets all trial information needed to generate a preview.

• TrialNumbers

public int TrialNumbers { get; set; }
Gets or sets the number of trials.

Constructors

• .ctor

public TrialData( )

BMPBuilder– TrialForm 147

D.1.11 Class TrialForm

This class is the GUI for entering trial information.

Declaration

public class TrialForm
: Form

Constructors

• .ctor

public TrialForm( )

• .ctor

public TrialForm( )

– Parameters
∗ experimentForm - The experiment form.

• .ctor

public TrialForm( )

– Parameters
∗ experimentForm - The experiment form.
∗ editNumber - The edit number of the trial to be changed.

List of Symbols
and Abbreviations

Abbreviation Description Deﬁnition
EyeTracker Eye tracking device (e.g. EyeLink II from SR Re- page 8
search)
FPOI Fixation pattern of interest page 31
R Radius page 53
ROI Region of interest page 30

149

List of Figures

2.1 EyeLink II eye tracking device . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 EyeLink II connection schema . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 EyeLink II camera setup screen shot . . . . . . . . . . . . . . . . . . . 9
2.4 Simple geometric target shapes, [14] . . . . . . . . . . . . . . . . . . . 10
2.5 Noise stimuli, with superimposed eye movements from a trial, [14] . 10
2.6 Statistically thresholded results for all subjects, [14] . . . . . . . . . . 11
2.7 Example of target pattern and search picture with 2 colors . . . . . . 13
2.8 Example of target pattern and search picture with 3 colors . . . . . . 14

3.1 Startup screen of Experiment Builder . . . . . . . . . . . . . . . . . . . 18
3.2 Window to enter trial block parameter . . . . . . . . . . . . . . . . . . 19
3.3 All necessary parameter information is entered . . . . . . . . . . . . . 21
3.4 A new block of trials has been added to the trial list . . . . . . . . . . 22
3.5 The File menu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Overview of all classes in the Experiment Builder . . . . . . . . . . . 24
3.7 Example for a target pattern . . . . . . . . . . . . . . . . . . . . . . . . 26
3.8 The search picture in which the target pattern has to be found . . . . 27
3.9 Target pattern and search picture on the same screen . . . . . . . . . . 28
3.10 Experiment Analysis Tool main GUI . . . . . . . . . . . . . . . . . . . 30
3.11 The experiment folder and data file have been selected . . . . . . . . 31
3.12 A queue was generated and the experiment analysis started . . . . . 34
3.13 Experiment Metadata . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.14 Trial information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.15 Details for four different fixation types, including ROI and target pat-
tern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.16 Color distribution example . . . . . . . . . . . . . . . . . . . . . . . . 38
3.17 Average vector of all FPOIs . . . . . . . . . . . . . . . . . . . . . . . . 38
3.18 Top 3 of a fixation ranking . . . . . . . . . . . . . . . . . . . . . . . . . 39

150


3.19 Example for the high similarity ranking . . . . . . . . . . . . . . . . . 40
3.20 Example for the overall sub-fixation ranking . . . . . . . . . . . . . . 41
3.21 Class diagram of ASCLibrary . . . . . . . . . . . . . . . . . . . . . . . 43
3.22 Class diagram of the Experiment Analysis Tool . . . . . . . . . . . . . 45

4.1 Patterns A1 to A4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Patterns B1 to B4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Patterns C1 to C4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4 Patterns D1 to D4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.5 Patterns E1 to E4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.6 Fixations superimposed in green, target pattern E3 in red. . . . . . . . 52
4.7 Pattern D1 fixation and high similarity ranking, radius 0 . . . . . . . 56
4.8 Pattern D3 fixation and high similarity ranking, radius 0 . . . . . . . 57
4.9 Pattern C3 fixation and high similarity ranking, radius 1 . . . . . . . 58
4.10 Pattern E4 fixation and high similarity ranking, radius 1 . . . . . . . . 59
4.11 Sub-fixation ranking for pattern A4, radius 1 . . . . . . . . . . . . . . 60
4.12 Patterns F1 and F2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.13 Patterns G1 and G2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.14 Patterns H1 and H2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.15 Patterns I1 and I2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.16 Patterns J1 and J2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.17 Patterns K1 and K2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.18 Patterns L1 and L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.19 Patterns M1 and M2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.20 Patterns N1 and N2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.21 Pattern G2 fixation and high similarity ranking, radius 1 . . . . . . . 70
4.22 Pattern M2 fixation and high similarity ranking, radius 1 . . . . . . . 71

List of Tables

4.1 Color distribution block 1 (Exp. I) . . . . . . . . . . . . . . . . . . . . . 53
4.2 Color distribution block 3 (Exp.I) . . . . . . . . . . . . . . . . . . . . . 53
4.3 Color distribution block 2 (Exp. I) . . . . . . . . . . . . . . . . . . . . . 54
4.4 Color distribution block 4 (Exp.I) . . . . . . . . . . . . . . . . . . . . . 54
4.5 Target pattern rankings in high similarity list, block 1 . . . . . . . . . 55
4.9 First place in sub-fixation rankings is part of target pattern, block 1
and 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.10 First place in sub-fixation rankings is part of target pattern, block 2
and 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.11 Target pattern rankings in high similarity list using duration as
weight, block 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
weight, block 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
weight, block 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
weight, block 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.15 Color distribution block 1 (Exp. 2) . . . . . . . . . . . . . . . . . . . . 68
4.16 Color distribution block 2 (Exp. 2) . . . . . . . . . . . . . . . . . . . . 68
4.19 First place in sub-fixation rankings is part of target pattern, block 1 . 71
4.20 First place in sub-fixation rankings is part of target pattern, block 2 . 71
weight, block 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

152


weight, block 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

Master Thesis - Algorithm for pattern recognition

Recommended

More Related Content

What's hot (18)

Viewers also liked (7)

Similar to Master Thesis - Algorithm for pattern recognition (20)

More from A. LE (9)

Recently uploaded (20)

Master Thesis - Algorithm for pattern recognition