An Experimental Analysis of Video Eye
in Predicting Human Engagement Behaviour
Tracker
Data
Tanmoy Debnath and Manoranjan Paul.
School of Computing and Mathematics, Charles Sturt University, Bathurst, NSW 2795, Australia.
Abstract
Fig. 5: Tobi X120 eye
tracker
has
been
employed in collecting
data (picture courtesy:
www.wired.com).
Fig. 1.: Participants’ eye gaze points for frame 1 of
different video clips as indicated by the colored dots.
The pupil sizes are measured in millimetre (mm), blinking
patterns in percentage (%) and all the human gaze distances are
.
measured in pixels in analysing the eye tracker extracted video
data. Both the pupil sizes (left and right) and fixations are
directly proportional to the EB, while any distance considered in
this experiment is inversely proportional. Hence the EB can be
expressed by the following function-

1 1 1 1 
EB = f  PS L , PS R , F , , , , 
DA DG DC DE 

Foreman
Hall monitor
Harbour
Mobile
54
49
47
56
51
38
31
39
48
41
49
43
64
50
32
58
42
46
28
40
56
51
49
30
51
49
51
50
37
35
31
52
44
46
43
57
58
49
55
61
55
36
21
38
46
48
41
38
58
61
46
50
54
35
29
36
55
38
37
47
41
50
53
56
47
30
29
40
60
48
33
59
Table 1: Overall ranking of
eighteen videos
Overall
Rank
Flower
43
55
39
45
45
40
36
49
54
48
39
53
Overall
Score
Crew
53
53
42
42
52
48
31
45
53
43
43
41
Waterfall
Container
54
34
54
51
59
41
42
42
58
31
43
37
Tempete
Coastguard
59
56
30
54
30
43
51
31
57
48
44
43
Stefan
City
46
58
48
56
38
34
47
60
30
45
47
37
Final Grades
Soccer
Bus
Fig. 2.: Average pupil sizes of 12 participants over 18
videos.
1
2
3
4
5
6
7
8
9
10
11
12
Akiyo




Left pupil size (PSL).
Right pupil size (PSR).
Blinking pattern (F).
Distance between individual’s average gaze location to all
participants’ average gaze location as obtained by the eye
tracker data (DA) .
 Distance between individual’s average gaze location to the
average
gaze location as predicted by the GBVS model
(DG).
 Distance between individual’s average gaze location to the
centre of
the videos (DC).
 Distance between individual’s average gaze location to the
most significant points according to the video experts (DE).
Video Clips and Corresponding Scores of Each Participants
Participants
Based on following seven descriptors, the performance of the
participants are graded in order to comprehend their
engagement behaviour (EB) with the videos:
Silent
Experimental Results and Analysis
Fig. 6: Video wise
average pixel, Experts’
opinion
and
GBVS
distances
from
the
videos’ centres.
Mother
daughter
News
In this work seven descriptors were utilized to rank twelve
participants’ engagement performance with eighteen common
intermediate format (CIF) videos which included experimental
results from an Tobi X120 eye tracker, video experts’ opinion,
and data from graph based visual saliency (GBVS) model- a well
known saliency modelling algorithm.
43
53
44
62
51
45
33
34
47
44
44
46
48
39
53
61
53
40
37
43
39
46
62
25
29
59
53
57
53
33
44
51
32
50
38
47
49
52
54
46
35
45
35
51
31
45
48
55
52
61
47
50
41
43
39
47
43
35
55
33
51
59
22
57
58
29
35
51
52
51
45
36
910
938
802
969
838
693
631
784
861
803
820
779
3rd
2nd
8th
1st
5th
11th
12th
9th
4th
57
51
32
57
37
32
32
35
56
45
60
52
7th
6th
10th
the twelve participants over
Future Work
Fig. 3: Video wise and person wise average fixation
percentage
At the moment equal weights are used for all the
descriptors. In future different descriptors would be
provided with different weights based on their relative
importance so that the human engagement behaviour
could be analyzed with greater efficiency.
Contact details:
Tanmoy Debnath
Fig. 4.: Video and person wise average pixel distances
from the videos’ centers.
Phone: +61 2 63384278
Email: [email protected]
www.csu.edu.au