Eye and Mouse Coordination During Task: From Behaviour to

Prediction

Alexandre Milisavljevic

1,2,3

, Kevin Hamard

, Coralie Petermann

, Bernard Gosselin

Karine Dor

e-Mazars

and Matei Mancas

Numediart institute, University of Mons, Mons, Belgium

Psychology institute, VAC EA7326 team, Paris Descartes University, Paris, France

Research and Development department, Sublime Skinz, Paris, France

Keywords:

Behaviour, Visual Attention, Webpages, Mouse-tracking.

Abstract:

The study of web users’ behaviour is of crucial importance for understanding people reaction when browsing

websites. Eye-tracking is a precise tool for this purpose, but it is hard to scale up when trying to apply it to a

wide range of situations and websites. On the other hand, mouse-tracking fulﬁlls these requirements. Unfor-

tunately, mouse data provides a limited approximation of the eye position as it was shown in the literature. In

this paper, we investigated the relationship between mouse and eye behaviour on several kind of websites with

three different tasks to create models based on these behaviours. Our ﬁndings were that 1) saliency Pearson’s

correlation is not suitable to analyse eye and mouse coordination, 2) this coordination is altered according to

the task, 3) scroll speed directly inﬂuence where the eyes are during the scroll, 4) amplitude vary according to

eyes position before the scroll and 5) by using the X axis variations it is possible to ﬁnd the moments where it

is easier to model eyes location from mouse location.

1 INTRODUCTION

Understanding why a user visits a webpage has been a

central question since the beginning of the twenty ﬁrst

century. To answer this question, Eye-tracking has

been used as a precise tool to estimate intention im-

pact on user’s gaze. However, these kinds of studies

are hard to scale up and apply it to a wide user panel

is difﬁcult. That is why mouse-tracking emerged as

an efﬁcient proxy to determine user’s attention. Since

then, correlation between mouse movements and eye

movements has been found (Mueller and Lockerd,

2001; Chen, 2001; Rodden and Fu, 2007; Rodden

et al., 2008; Cooke, 2006; Guo and Agichtein, 2010;

Huang and White, 2012; Navalpakkam et al., 2013)

and modelling attempts followed (Guo and Agichtein,

2010; Huang and White, 2012; Navalpakkam et al.,

2013; Boi et al., 2016).

Nevertheless, a majority of these studies focused

on SERP (SEarch Result Pages) from search engines

putting aside the rest of the web and tasks. In addi-

tion, eye-mouse and eye-task relationships have been

studied separately (Yarbus, 1967; Castelhano et al.,

2009; Mills et al., 2011) but rarely together. That is

why, the goal of this study was to explore the eye-

mouse-task relationship in a more diversiﬁed environ-

ment.

Chen (2001) was the ﬁrst to show that areas visi-

ted by the mouse were also visited by the eye in free-

viewing condition. Rodden and Fu (2007) also sho-

wed that regions visited by the mouse were also visi-

ted by the eye but they were the ﬁrst to highlight the

better correspondence on Y axis between mouse and

eye. Unlike previous work, they set-up an experiment

with pre-deﬁned search queries on a search engine.

Guo and Agichtein (2010) conﬁrmed Rodden and Fu

(2007) results about more accurate correlation on Y

axis. Their main contribution was the ﬁrst attempt

to automatically infer the user’s eye position using

mouse movements. They also suggested the presence

of images did not have a signiﬁcant effect on eye-

mouse coordination. Huang and White (2012) presen-

ted that amount of time spent on a search web page by

a participant can affect where they were pointing and

looking and then used this ﬁnding to enhance their

algorithm. They showed that gaze-cursor alignment

was distinct for each participant but did not highlight

signiﬁcant difference among women and men. Naval-

pakkam et al. (2013) updated previous work by inves-

tigating more recent SERP which now includes ima-

ges and more complex content. They showed that this

content induced different behaviour. Then they pro-

posed a non-linear model outperforming state-of-the-

Milisavljevic, A., Hamard, K., Petermann, C., Gosselin, B., Doré-Mazars, K. and Mancas, M.

Eye and Mouse Coordination During Task: From Behaviour to Prediction.

DOI: 10.5220/0006618800860093

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 2: HUCAPP, pages

86-93

ISBN: 978-989-758-288-2

Table 1: Generic tasks used in the study.

Category Description

Free viewing Browse the website by visiting at

least two other pages

Browse two articles of your

choice

Target ﬁnding Browse the following pages: ca-

lendar, team and news

Buy the speciﬁc given item

Text reading Read the two ﬁrst paragraphs

art models because of its non-linearity.

Our exploratory study aimed to investigate ef-

fect(s) of user’s goal on eye and mouse coordination

in ecological conditions (different categories of web

sites, no scroll limitations). Our hypothesis was the

following: there is a direct link between the eye, the

mouse movements and the task which ﬂuctuates ac-

cording to the task. Thus, we could enhance the pre-

cision of the current models.

This paper was structured as follows: experiment

set-up was described in section 2 followed by the re-

sults of the static and dynamic analyses in section

3. Finally we discussed and concluded the paper in

section 4.

2 METHOD

We recruited ﬁve participants with normal or

corrected-to-normal vision (4 males and 1 female)

aged between 24 and 25 years from the local signal

processing department. All participants were right-

handed and ﬂuent with computer operations. They

were tested on 10 various websites including blogs, e-

commerce platforms, etc. From the calibration phase

at the beginning of the study to the end, the whole

process took about 20 minutes per person.

2.1 Tasks

We used a set of ﬁve tasks distributed in three classes

as presented in Table 1: free viewing, target ﬁnding

and text reading. The number of pages that could

be visited during Free viewing tasks was limited to

two but was not limited in time and participants had

full scroll possibilities. The reading task was speci-

ﬁc enough to prevent any free interpretation in order

to simulate participants’ willingness to read a speciﬁc

paragraph. Finally, we chose two types of target ﬁn-

ding tasks: one in which participants were instructed

to ﬁnd and buy an item, the second in which they had

to ﬁnd a given page.

2.2 Set-up

To record eye movements, we used a FaceLAB 5 eye-

tracker at 60Hz without head constraint on a 17-inch

screen set to a resolution of 1920X1080. Instructions

and websites were displayed in Google Chrome max-

imized with a resolution of 1920X955.

To record Mouse movements we developed a

plug-in using WebExtensions

standard. It took the

form of an ON/OFF button on the browser top bar and

has only been used by the operator. The extension

monitored the following metrics: time-stamp, event

type (click, movement or scroll), mouse’s coordina-

tes, offset induced by the scroll, URL and screen size.

The plug-in developed in Javascript was uploading all

mentioned metrics on the ﬂy or every 50-60ms to a

NodeJS

server via a socket connection. The same

server inserted in real time the data in a MySQL data-

base without further processing. The server also kept

track of the page to deliver to the participant.

2.3 Procedure

Participants started on a homepage describing the

context of the study. To visit the next planned page

by the study they had to click on a Javascript book-

mark situated in the browser’s bookmark top bar. The

ﬁrst click on it led them to the ﬁrst task instruction.

All tasks were stored in HTML format locally. After

reading the instruction, participants could once again

click on the bookmark and begin the task. When the

task was completed participants had to click again on

the bookmark to read the next instruction and so on.

At the end of the study, participants were asked to

answer to a survey about there knowledge of the web-

sites.

3 RESULTS

We ran three sets of analyses in order to highlight

coordination between eye and mouse movements.

First, we used 2D saliency metric PCC (Pearson’s

Correlation Coefﬁcient) to check consistency bet-

ween overall eye and mouse movements. Then we

repeated the same analysis between participants’ eyes

movements. Second, we applied literature’s tempo-

ral and distance estimation to our task-related context

to bring out tasks’ inﬂuence on eye and mouse coor-

dination. Third, we analysed the participants’ beha-

viour while scrolling because - at our knowledge - it

https://developer.mozilla.org/Add-ons/WebExtensions

https://nodejs.org/en/

Eye and Mouse Coordination During Task: From Behaviour to Prediction

Figure 1: (a) eye ﬁxation density map, (b) original website and (c) mouse ﬁxation density.

has not been treated by the literature whereas it could

be an essential information to the understanding of

eye and mouse coordination. Finally, we used the re-

sults of the two ﬁrst sets of observations to create two

Gaussian-based models to approximate eye position

from mouse position.

While the ﬁrst approach focused on a static and

spatial analysis, the second and third aimed for a dy-

namic analysis taking into account the temporal evo-

lution of both eye and mouse tracks.

3.1 Static Fixation Densities

Comparison

Pearson’s Correlation Coefﬁcient (PCC) also known

as the Pearson Product-Moment Correlation (1), is a

metric used in saliency maps comparison by authors

like Ouerhani et al. (2004) and Le Meur et al. (2007)

and used to compare ﬁxations and mouse movements

by Tavakoli et al. (2017). PCC has a value between

-1 and 1. When the coefﬁcient is almost equal to 1,

there is a strong relationship between the two varia-

bles. The goal was to apply this metric to highlight

eye and mouse coordination changes between tasks.

The originality of this metric lies in the fact that it uses

probability densities instead of raw variables values.

To do so, we computed PCC between eyes density

map and mouse density map. To obtain these maps,

ﬁxations from eye-tracking and mouse-tracking were

convolved with a Gaussian ﬁlter. Thus PCC was com-

puted between images (a) and (c) as shown in Figure

Table 2: Pearson’s correlation coefﬁcients for intra and inter

analyses.

Task inter intra-eye intra-mouse

Free viewing 0.132 0.036 0.082

Target ﬁnding 0.171 0.028 0.107

Text reading 0.176 0.440 0.162

X,Y

cov(X,Y )

(1)

We obtained for our three tasks classes (free vie-

wing, target ﬁnding and text reading) deﬁned in

section 2.1 correlation scores as in Table 2, “inter”

column. Both classes correlation and their relative

difference remained small which showed that mouse-

tracking could not be directly used to model eye mo-

vements. For this reason, we decided to reﬁne the

investigation based on motion dynamics in the next

sections.

Furthermore, when comparing eye-tracking re-

sults between different participants on the same sti-

mulus, we obtained results in Table 2, “intra-eye” co-

lumn which showed a higher correlation for “Text re-

ading” task than for the two others. This result conﬁr-

med that if the task and its location were precise, then

most of the participants would produce similar eye-

gaze patterns. We observed the same behaviour for

mouse tracks in Table 2, “intra-mouse” column, but

with a lower overall correlation which showed that

mouse behaviour remained less consistent than eye

behaviour.

HUCAPP 2018 - International Conference on Human Computer Interaction Theory and Applications

Figure 2: First column (a) is a free viewing task, second column (b) is a target ﬁnding task and third column (c) is a text

reading task.

3.2 Dynamic Analyses

Considering the dominant use of scroll in our expe-

riment, modern vertically-based designs and the tend

of Human eye to be more efﬁcient horizontally, we

separated X and Y coordinates to enhance granularity

in our dynamic analyses. For each X and Y coordinate

we got temporal vectors which were synchronized be-

tween mouse and eye. To do so, we matched eye

ﬁxations with mouse events and then down-sampled

mouse data to ﬁt eye data. We chose to not interpo-

late as in Deng et al. (2016) because it could have

generated non-existing ﬁxations and wrong results.

3.2.1 Temporal and Distance Estimation

We observed for some participants a time shift on Y

axis between mouse and eye with the mouse being

delayed as in Figure 2 (a) and (c) right columns.

This ﬁnding joined Huang and White (2012) previous

work in which they detected a lag between the mouse

and the eye. This could be explained by the fact that,

in visual exploration context, the eye is the only mean

of perception and leads the hand movements.

We computed euclidean distance (3) and obtai-

ned an eye-mouse distance of 554 pixels. This re-

sult was not in accordance with the average 229 pixels

from state-of-the-art (Rodden and Fu, 2007; Guo and

Agichtein, 2010; Huang and White, 2012). We then

reﬁned our analysis by separating the two axes. Using

formula (2) we got a mean distance of 409 pixels for

X axis and 291 pixels for Y axis. With this results

we began to have a better consistency on Y axis as

expected. However, Bejan (2009) demonstrated that

our eyes scan horizontally faster than in the vertical

dimension. Based on our results, we could assumed

that participants kept their mouse vertically stationary

to scroll down or up and used it as a vertical pointer,

allowing them to horizontally browse without difﬁ-

culties. Thus the participant could easily move his

eye on X axis more often. That is why the participant

tended to move it’s eye on X axis more often.

We then continued with separate axes to compute

correlation. Compared to distances, correlation coef-

ﬁcients between mouse and eye were drastically diffe-

rent. Chen (2001) obtained a correlation of 0.58 with

more than 50% of the pages associated with correlati-

ons larger than 0.8. In our study, we measured a mean

correlation of 0.64 on Y axis and 0.18 on X axis.

Difference between axes got even more signiﬁcant

when we examined these correlations coefﬁcients ac-

cording to their corresponding task. As exposed in

Table 3, free viewing task had the best correlation on

Y with 0.9. This result reﬂected a greater trend to use

the mouse as a vertical pointer as in other tasks. Coef-

ﬁcients for target ﬁnding were more balanced with an

increased correlation on X but a decrease on Y. Fi-

nally, text reading correlations expressed the fact that

participants did not used much the mouse during this

task. We could assume that more the cognitive load

of the task is important more the correlation drop on

both axes.

Eye and Mouse Coordination During Task: From Behaviour to Prediction

Table 3: Pearson’s correlation on X and Y axes.

Task type r

Free viewing 0.176 0.921

Target ﬁnding 0.383 0.699

Text reading 0.006 0.32

d(i) = |x

(i) − x

(i)| (2)

d(X,Y ) =

− x

)

+ (y

− y

)

(3)

3.2.2 Scroll’s Speed and Direction Inﬂuence

As we previously exposed, mouse and eyes were more

correlated on Y axis. In addition, scroll events are a

barely studied subject while it is a common behaviour

in all webpages browsing. We based all our calculus

on scroll sessions which corresponds to a set of con-

tinuous scroll events ended with a mouse movement.

Scroll is an important feature providing good infor-

mation about the degree of participants’ interest on a

website. Another advantage is that the scroll is me-

asurable on desktop and mobile. Through the follo-

wing analyses, we highlighted inﬂuence of behaviour

on scroll’s speed and amplitude.

We collected for each scroll session the direction (up

or down) and the absolute speed. After empirical tries

and errors and after taking into account the amount

of data, we also separated the browser screen into 3

equals categories as in Figure 5 to detect patterns.

For the current analyses we removed text reading

tasks because it did not included enough scroll events.

For both amplitude and speed inﬂuence test, we per-

formed a one-way independent ANOVA (analysis of

variance) (4) test. The ANOVA examines if the mean

of numeric variables differs across levels of catego-

rical variables. After checking all assumptions (nor-

mality of errors, equal error variance across category,

independence of errors), we hypothesized:

: µ

= µ

(4)

: At least one mean is not equal to the others.

As shown in Table 4, we considered that all means

were equal to each other. The statistic test we ran

was the ratio of the between-category variance and

the within-category variance. If this ratio was gre-

ater than the critical probability distribution F, we

could reject the null hypothesis. After obtaining a p-

value below the 0.05 threshold, we could afﬁrm the

rejection of the null hypothesis with a conﬁdence rate

of 95 %. Thus, we can conclude that there is an effect

of scroll speed on eyes category position.

Table 4: Result test ANOVA with signiﬁcance level (p-

value) and F-score.

Indicator Task Down Up

F-test Free viewing 4.26 7.07

Target ﬁnding 3.76 -

P-value Free viewing 0.017 0.001

Target ﬁnding 0.031 -

To go further, we had to determine and deﬁne this

inﬂuence. We focused on means for each tasks using

a Tuckey’s test. We observed that while scrolling

quickly, participants positioned their eyes at the

opposite side of the scroll’s direction to be able to

detect bottom-up characteristics through peripheral

vision as shown in Figure 3. Furthermore, when

participants were looking for a speciﬁc information

(target ﬁnding task), they tended to quickly look

towards the center of the screen when the scroll speed

decreased.

Then we focused on scroll’s amplitude, which is

the distance between the start and the end of a scroll

session. We wanted to know if participants adapted

their eyes position before scrolling. Here again we

differentiated the target ﬁnding and the free viewing

tasks and calculated the means distance for each area

before the participant scroll.

Table 5: Result test ANOVA with signiﬁcance level (p-

value) and F-score.

Indicator Task Down Up

F-test Free viewing 3.08 10.44

Target ﬁnding 0.09 -

P-value Free viewing < 0.001 < 0.001

Target ﬁnding 0.9 -

We could conclude using ANOVA test that for

the free viewing task, when the eyes were located

at the bottom of the screen and before scrolling up,

the scroll amplitude increased with p-value < 0.05 as

shown in Table 5. As expected, when the eyes were

located on the top of the screen before scrolling down

the scroll amplitude were much higher, see Figure 4

(a) and (b). About the target ﬁnding task, there was

no signiﬁcant impact (c) of amplitude on the screen

area before scrolling (p-value > 0.05). However, we

noticed that when searching speciﬁc information, par-

ticipants did not have a long scroll amplitude in order

to not miss an element (text, blocs, titles, etc) and to

differentiate them.

The scroll event could improve the prediction of

the localization of the eyes on Y axis using the com-

bination of direction, amplitude and speed variables.

HUCAPP 2018 - International Conference on Human Computer Interaction Theory and Applications

Figure 3: Eyes position according to scroll speed, (a) and (b) corresponds to free viewing task, (c) corresponds to target

ﬁnding task.

Figure 4: Scroll amplitude according to screen area before scroll, (a) and (b) corresponds to free viewing task, (c) corresponds

to target ﬁnding task.

Figure 5: Screen’s three areas.

3.3 Model

Previous analyses provided several insights about

users behaviours on webpages given more or less spe-

ciﬁc tasks. We built our models from these, more par-

ticularly from the eyes movements standard deviati-

ons. As in section 3.2.1, we separated X and Y axes to

infer the parameters of a Gaussian model which pre-

dicted the eyes position based on the mouse position

and cognitive load of the task. From these standard

deviations we were able to deﬁne a conﬁdence area

around the mouse in which the eyes had a 70% pro-

bability to be in it. We chose to base our calculus on

the 70

percentile because it was the minimum con-

ﬁdence rate we observed in the state of the art. As

shown in Table 6, columns “x std.” and “y std.”, the

percentile (5) gave a ﬁrst coarse pixel area around

the mouse cursor.

percentile = µ ± Zσ (5)

Table 6: Standard deviation (percentile 70%) normal and

using only sudden X changes.

Task x std. y std. x std.

thrs.

y std.

thrs.

Free

viewing

558.0 416.4 - -

Target

ﬁnding

486.4 403.8 361.3 251.0

Text re-

ading

627.8 257.9 - -

But we were interested in a better model, so we focu-

sed on speciﬁc behaviours during tasks. As shown in

Figure 2 (a) target ﬁnding class, we identiﬁed sudden

changes on X axis. After analysing participants’ vi-

deos and comparing with several target ﬁnding tasks

among them, we found that these sudden changes ma-

tched participant’s interest. When the participant had

a target ﬁnding goal and found his target, he quickly

moved his mouse to the point of interest.

Thus, we manually deﬁned a threshold at the be-

ginning of each sudden changes and we computed the

standard deviation before and after every ﬁnal sud-

den change on X. We obtained better results as shown

in Table 6, column “x std. thrs.” and “x std. thrs.”.

Area covered by the 70

was reduced by around 150

pixels for both axes. With this second model, we were

able to increase the accuracy but only by focusing on

a speciﬁc event.

Eye and Mouse Coordination During Task: From Behaviour to Prediction

4 CONCLUSION AND

DISCUSSION

We ﬁrst compared eye and mouse data with the sa-

liency metric PCC. We did not ﬁnd signiﬁcant con-

sistency between participants’ eyes and mouse positi-

ons (inter) and between participants’ eyes (intra). Ho-

wever, results showed that participants behaved in a

more similar way when they had the same task with

the same location (reading task).

Then, we got deeper with dynamic analyses. We

showed that using distance and correlation, we were

able to highlight more interesting coordinations bet-

ween eyes and mouse. We had better results on Y axis

than X axis and succeed to demonstrate behaviour dif-

ferences between tasks. In addition, scroll analyses,

clearly showed a relation between eyes position and

scroll speed while browsing and amplitude before the

scroll.

Finally, we made a model for each task able to

predict the area around the mouse’s cursor in which

the eyes had 70% chances to be located in. Howe-

ver, eyes location uncertainty compared to mouse po-

sition remained high, even if we succeed to enhance

the model during target ﬁnding task by observing bru-

tal changes on X axis.

In this paper, we presented results of a prelimi-

nary study, used as a validation to conduct a bigger

experiment, including more participants. This will al-

low us to analyse the impact of participants’ age on

their mouse movements. Moreover, we did not use

scroll events analyses to enhance our models. In fu-

ture work, we think that doing so, could boost the

precision of the model by reducing the area around

the mouse’s cursor. We could also investigate new re-

lations between the scroll and the eyes by analysing

scroll in 2D. Then, we could use machine learning

models to integrate new features and more user beha-

viours such as mouse patterns. Finally, our main ob-

jective is to propose the most accurate model in order

to use it in real time to predict web user behaviours.

ACKNOWLEDGEMENT

We thank French Research and Technology Associa-

tion (ANRT) and Sublime Skinz for supporting this

work.

REFERENCES

Bejan, A. (2009). The goldenratio predicted: Vision, cog-

nition and locomotion as a single design in nature. In-

ternational Journal of Design & Nature and Ecodyn-

amics, 4(2):97–104.

Boi, P., Fenu, G., Davide Spano, L., and Vargiu, V. (2016).

Reconstructing User’s Attention on the Web through

Mouse Movements and Perception-Based Content

Identiﬁcation. Transactions on Applied Perception,

13(3):1–21.

Castelhano, M. S., Mack, M. L., and Henderson, J. M.

(2009). Viewing task inﬂuences eye movement cont-

rol during active scene perception. Journal of Vision,

9(3):1–15.

Chen, M.-c. (2001). What can a mouse cursor tell us more?

Correlation of eye / mouse movements on web brow-

sing. In Conference on Human Factors in Computing

Systems, pages 281–282.

Cooke, L. (2006). Is the Mouse a” Poor Man’s Eye Trac-

ker”? Annual Conference-Society for Technical Com-

munication, 53:252 – 255.

Deng, S., Chang, J., Kirkby, J. A., and Zhang, J. J. (2016).

Gaze–mouse coordinated movements and dependency

with coordination demands in tracing. Behaviour &

Information Technology, 35(8):665–679.

Guo, Q. and Agichtein, E. (2010). Towards predicting web

searcher gaze position from mouse movements. In Ex-

tended Abstracts of Conference on Human Factors in

Computing Systems, pages 3601–3606.

Huang, J. and White, R. (2012). User See, User Point: Gaze

and Cursor Alignment in Web Search. In Special Inte-

rest Group of Conference on Human Factors in Com-

puting Systems, pages 1341–1350.

Le Meur, O., Le Callet, P., and Barba, D. (2007). Predicting

visual ﬁxations on video based on low-level visual fe-

atures. Vision Research, 47(19):2483–2498.

Mills, M., Hollingworth, A., and Dodd, M. D. (2011). Ex-

amining the inﬂuence of task set on eye movements

and ﬁ xations. Journal of Vision, 11(8):1–15.

Mueller, F. and Lockerd, A. (2001). Cheese: tracking

mouse movement activity on websites, a tool for user

modeling. In Extended Abstracts of Conference on

Human Factors in Computing Systems, pages 279–

280.

Navalpakkam, V., Jentzsch, L. L., Sayres, R., Ravi, S.,

Ahmed, A., and Smola, A. (2013). Measurement and

modeling of eye-mouse behavior in the presence of

nonlinear page layouts. In International Conference

on World Wide Web, pages 953–964.

Ouerhani, N., Wartburg, R. V., and Heinz, H. (2004). Empi-

rical Validation of the Saliency-based Model of Visual

Attention. Electronic Letters on Computer Vision and

Image Analysis, 3(1):13–24.

Rodden, K. and Fu, X. (2007). Exploring how mouse mo-

vements relate to eye movements on web search re-

sults pages. In Special Interest Group on Information

Retrieval Workshop on Web Information Seeking and

Interaction, pages 29–32.

Rodden, K., Fu, X., Aula, A., and Spiro, I. (2008). Eye-

mouse coordination patterns on web search results pa-

ges. In Extended Abstracts of Conference on Human

Factors in Computing Systems, pages 2997–3002.

HUCAPP 2018 - International Conference on Human Computer Interaction Theory and Applications

Tavakoli, H. R., Ahmed, F., Borji, A., and Laaksonen, J.

(2017). Saliency Revisited: Analysis of Mouse Mo-

vements versus Fixations. In Conference on Computer

Vision and Pattern Recognition, pages 4321–4329.

Yarbus, A. L. (1967). Eye Movements. New York: Plenum

Press.

Eye and Mouse Coordination During Task: From Behaviour to Prediction