DoYou See What I See?
The Effect of GazeTracking
onTask Space Remote Collaboration
Kunal	Gupta1,	Gun	A.	Lee2,	and	Mark	Billinghurst2	
1HIT	Lab	NZ,	University	of	Canterbury	
2Empathic	CompuIng	Lab,	University	of	South	Australia	
	
ISMAR	2016,	September	20th,	2016
Mo#va#on
• Improving	remote	assistance	of	expert	user	(e.g.	maintenance)	
• SupporIng	rich	communicaIon	cues
Task Space Teleconferencing
• Focus	on	sharing	view	of	remote	task	space	
• Methods	
•  Handheld	tablets	with	cameras	+	AR	cues,	Fixed	cameras	in	workspace	
• LimitaIons	
•  Fixed	viewpoint,	Difficult	to	know	where	user	looking
Head Worn Collabora#ve Systems
• Place	camera	on	head	+	use	head	mounted	display	
•  HWC	+	HMD	+	remote	poinIng	improves	collaboraIon	
• LimitaIons	
•  Remote	view	fixed,	Expert	doesn’t	know	exactly	where	worker	looking
Gaze Tracking in Teleconferencing
•  Monitor	based	(Brennan	2008),	(Carle[a	2010)	
•  Gaze	provides	a[enIon	cue,	significantly	improved	performance	
•  Head	mounted	(Fussell	2003),	(Ou	2005)	–	no	HMD	
•  No	performance	improvement,	focus	of	a[enIon	can	be	predicted
Comparison to Previous Work
Rem	=	remote	collaboraIon,	FtF	=	face	to	face	collaboraIon	
Gaze	=	eye	tracking,	HWC	=	head	worn	camera,	HMD	=	head	mounted	display
Key Research Ques#ons
• Q1:	Will	sharing	of	gaze	and	pointer	cues	affect	the	
feeling	of	co-presence	between	users?	
• Q2:	Will	sharing	of	gaze	and	pointer	cues	improve	
performance	in	a	remote	collaboraIve	task?
Prototype Design
• Combining	the	following	
(1)	a	head	mounted	eye-tracker	
(2)	head	mounted	camera	
(3)	head	mounted	display	
(4)	remote	viewing	sohware	
System	Diagram
Local Worker
Brother	Air	Scouter	
Microsoh		
Lifecam	
HD	5000	
Logitech	Webcam	c920
Pupil Labs Eye Tracking
• Open	source	eye-tracking	
• Use	IR	reflecIon	into	eye	
• Image	processing	on	PC	
• Tracks	eye	at	30	fps	
• Provides	raw	data	
• www.pupil-labs.com
Remote Expert Desktop
• Live	camera	view	
• Gaze	shown	as	red	dot	
• Can	add	pointer	cues	
•  Mouse	input	
•  Green	dot	
• Shown	in	HMD
User Experiment
• In remote expert collabora#on ... 
• Does Pointer / Eye tracker cues have significant effect
on co-presence?
• Does Pointer / Eye tracker cues have significant effect
on task performance?
Experimental Design
Eye tracker
cue
Pointer cue
No Yes
No NONE E
Yes P Both
Experimental Design – Setup
Experimental Design – Task
• Block	assembly	
• Four	different	structures	
• 17	pieces	in	each	
• Pilot	tested	to	balance		
difficulty	level	
• Assigned	to	condiIons	with	counter	balancing	
• AcIve	head	movement	encouraged	through	
secondary	task	(Imer)	and	L-shape	desk	setup
Experimental Design – Procedure
• PracIce	trial	in	face-to-face	collaboraIon	
• ParIcipants	separated	for	the	experimental	trials	
• For	each	condiIon:	
• Remote	helper	creates	structure	based	on	instrucIon	
• Perform	experimental	task	
• Answer	per-condiIon	quesIonnaire	
• Post-experimental	quesIonnaire	&	debriefing
Experimental Design – Par#cipants
• Within-subject	
• Balanced	LaIn	square	design	
• 30	parIcipants	(15	pairs)	recruited,	26	retained	
• 21-33	years	old,	73%	male	
• Fluent	English	speaking	
• No	one	had	done	block	assembly	over	video	
conferencing	before
Results - Summary
•  Both	the	POINTER	and	EYE	TRACKING	visual	cues	helped	parIcipants	to	perform	
the	task	significantly	faster.	
•  The	POINTER	cue	significantly	improved	both	local	and	remote	users’	perceived	
quality	of	communicaIon,	collaboraIon,	and	co-presence.
•  The	EYE	TRACKING	significantly	improved	the	communicaIon	and	collaboraIon	
quality,	and	sense	of	being	focused	for	local	workers,	and	enjoyment	for	remote	
helpers.
•  The	BOTH	condiIon	ranked	as	the	best	in	most	of	the	aspects	of	user	experience,	
while	the	NONE	condiIon	was	ranked	as	the	worst.
•  Visual	cues	made	the	conversaIon	more	efficient,	changed	the	choice	of	wording	
in	deicIc	expressions,	and	helped	parIcipants’	feel	more	connected.
Results – Task comple#on #me
• Repeated	measure		
two-way	ANOVA	(α	=	0.05)	
• POINTER	cue	
• F	(1,12)=4.908,	p=.047*	
• 15%	less	Ime 		
• EYETRACKER	cue	
• F	(1,12)=5.811,	p=.033*	
• 10%	less	Ime	
• InteracIon	
• F	(1,12)=0.566,	p=.466
sec.
Results – Per-condi#on ra#ng ques#onnaire
• Q1 	I	felt	connected	with	my	partner.	
• Q2 	I	felt	I	was	present	with	my	partner.	
• Q3 	My	partner	was	able	to	sense	my	presence.	
• Q4 	My	partner	(or	for	Remote	Helper:	I)	could	tell		
	when	I	(or	for	Remote	Helper:	my	partner)	needed	assistance.	
• Q5 	I	enjoyed	the	experience.	
• Q6 	I	was	able	to	focus	on	the	task	acIvity.	
• Q7 	I	am	confident	that	we	completed	the	task	correctly.	
• Q8 	My	partner	and	I	worked	together	well.	
• Q9 	I	was	able	to	express	myself	clearly.	
• Q10	I	was	able	to	understand	partner’s	message.	
• Q11	Informa9on	from	partner	was	helpful.	
Adopted from [Kim et al. 2014]
Results – Per-condi#on ra#ng ques#onnaire
• 7-point	Likert	Scale	
• 1:		totally	disagree	~	7:		totally	agree	
• Internal	consistency:		Cronbach’s	α=.937	
• Aggregated	into	0~100	scale
Results – Per-condi#on ra#ng ques#onnaire
• Aligned	Rank	Transform	
(ART)	+	Repeated	measure	
ANOVA	
	(α	=	0.05)	
• POINTER	cue	
• F	(1,12)=7.414,	p=.019*		
• EYETRACKER	cue	
• F	(1,12)=26.822,	p<.001*	
• InteracIon	
• F	(1,12)=2.023,	p=.180
Results – Per-condi#on ra#ng ques#onnaire
• Aligned	Rank	Transform	(ART)	
+	Repeated	measure	ANOVA	
	(α	=	0.05)	
• POINTER	cue	
•  F	(1,12)=11.914,	p=.005*	
• EYETRACKER	cue	
•  F	(1,12)=15.929,	p=.002*	
• InteracIon	
•  F	(1,12)=5.157,	p=.042*
Results – Per-condi#on ra#ng ques#onnaire
• Local	Workers
Results – Per-condi#on ra#ng ques#onnaire
• Remote	Helpers
Results – Ranking
Results – Ranking
• Friedman	test		
(α	=	0.05)	
• Local	worker	
• E	>	None	on	C5	
• Remote	helper	
• E	>	None	on	C2,4,5	
1: the best ~ 4: the worst
Results – Preference and qualita#ve feedback
• Understanding	partner	
•  Local	workers:	85%	preferred	condiIons	including	POINTER	cue	
“With	Pointer,	I	can	relate	to	what	he	is	talking	about,	because	I	could	understand	
him	more.”	
•  Remote	helpers:	70%	preferred	the	BOTH	condiIon	
“The	eye	tracker	helps	me	to	look	in	the	same	view	of	my	partner.”	
• Performing	task	efficiently	
•  77%	of	Local	&	85%	of	Remote	users	preferred	the	BOTH	condiIon	
“The	eye	tracker	was	giving	my	partner	more	informaFon	about	where	I	looked	at,	
while	the	pointer	was	for	giving	me	the	instrucFon	from	my	partner.”
Results – Behaviour observa#on
• Pointer	cue	reduced	number	of	phrases	said	
•  Local	worker	F	(1,11)=6.532,	p=.027*	
•  Remote	helper	F	(1,11)=8.479,	p=.014*	
• Referring	to	objects	&	direcIng	
Without	Pointer With	Pointer
describe	features		
(colour,	size,	shape,	...)
“this	one”
“move	leG/right”,	“in	front	of	” “put	it	here”,	“next	to	this”
“that”	object “this”	object
Discussion
• Performance	improved	by	using	eye-tracker	and	pointer	
•  Pointer	–	giving	direct	guidance	
•  Eye-tracker	–	showed	workers	focus	of	a[enIon	
• Eye	tracker	provided	benefit	even	without	pointer	
•  Same	view	condiIon	for	local	worker,	but	improved	communicaIon	quality	
• Benefit	of	gaze	informaIon	
•  Improved	communicaIon,	more	enjoyable,	focus,	reduce	interrupIon	
• Benefit	of	virtual	pointer	cue	
•  Ease	of	direcIon,	increased	sense	of	presence,	increased	partner	awareness	
• Different	user	needs	
•  Local	worker	–	understanding/empathy	–	benefits	from	eye-tracking	
•  Remote	expert	–	giving	instrucIon	–	benefits	from	using	pointer	cue
Implica#ons 
1.  Eye-tracking	can	be	used	to	change	the	nature	of	remote	
collaboraIon	with	head	worn	systems	
•  Make	remote	user	aware	of	implicit	intenIons	
2.  Providing	gaze	cues	alone	can	significantly	improve	the	remote	
collaboraIon	even	without	remote	poinIng	
•  eye-tracking	just	as	beneficial	as	using	remote	poinIng	by	itself	
3.  CommunicaIon	cues	like	gaze	and	poinIng	play	a	very	important	role	
in	creaIng	a	sense	of	co-presence	and	deeper	understanding	
•  Most	of	the	users	preferred	gaze	+	pointer	due	to	connecIon	created
Limita#ons 
1.  Prototype	too	bulky	and	tethered	to	PC,	not	wearable	
• New	HMDs	with	integrated	cameras	could	overcome	this		
2.  Task	limited	compared	to	realisIc	remote	collaboraIon	
• Did	have	key	elements	such	as	object	idenIficaIon	
3.  Experimental	measures	not	so	detailed	
• Detailed	behavioral	analysis,	conversaIonal	analysis
Conclusions
• Does	sharing	eye	tracking	informaIon	between	local	user	and	
remote	expert	help	in	terms	of	co-presence	and	performance?	
•  Using	gaze	and	pointer	a[enIon	cues	improved	performance	Ime.	
•  Gaze	and	pointer	cues	improved	the	feeling	of	co-presence		
• Many	areas	for	future	work	
•  Explore	parallel	task	–	both	people	with	same	roles	
•  Provide	symmetric	communicaIon	cues	–	gaze	both	ways	
•  Use	other	physiological	cues,	GSR,	heart	rate,	EEG
Empathy Glasses (CHI 2016)
• Combine	together	eye-tracking,	display,	face	expression	
• Implicit	cues	–	eye	gaze,	face	expression	
+	+	
Pupil	Labs	 Epson	BT-200	 AffecIveWear	
					
Masai,	K.,	Sugimoto,	M.,	Kunze,	K.,	&	Billinghurst,	M.	(2016,	May).	Empathy	Glasses.	In	Proceedings	of	
the	34th	Annual	ACM	Conference	Extended	Abstracts	on	Human	Factors	in	CompuFng	Systems.	ACM.
Affec#veWear – Emo#on Glasses
• Photo	sensors	to	recognize	expression	
• User	calibraIon	
• Machine	learning	
• Recognizing	8	face	expressions
Remote Collabora#on
• Eye	gaze	pointer	and	remote	poinIng	
• Face	expression	display	
• Implicit	cues	for	remote	collaboraIon
Contact Us
www.empathiccomputing.org
@marknb00
mark.billinghurst@unisa.edu.au
gun.lee@unisa.edu.au

Ismar 2016 Presentation