55
T
elema
tics
Effectiveness analysis of selected attention models
Model (method)
β
Std. Error
t value
Pr(>|t|)
SUN ICA (3)
Constant
398.350576
3.524468
113.024309
0.000001
SOA
-0.054045
0.007637
-7.076492
0.000001
Saliency
-0.310456
4.972281
-0.062437
0.950218
SUN DOG (3)
Constant
392.503239
3.800210
103.284630
0.000001
SOA
-0.053838
0.007633
-7.053486
0.000001
Saliency
11.072075
5.034701
2.199152
0.027920 ***
Bruce & Tsotos (3)
Constant
394.591120
3.683474
107.124731
0.000001
SOA
-0.053943
0.007635
-7.065408
0.000001
Saliency
7.496456
5.010818
1.496054
0.134713
Itti, Koch & Niebur (3)
Constant
396.595948
3.234775
122.603870
0.000001
SOA
-0.054006
0.007636
-7.072695
0.000001
Saliency
4.675457
4.798386
0.974381
0.329922
Zhai & Shah (4)
Constant
393.760655
3.537627
111.306434
0.000001
SOA
-0.053868
0.007633
-7.056902
0.000001
Saliency
10.270012
5.060635
2.029392
0.042480 ***
SUN ICA (4)
Constant
397.833645
3.511731
113.287039
0.000001
SOA
-0.054019
0.007637
-7.073006
0.000001
Saliency
0.887398
5.006221
0.177259
0.859313
SUN DOG (4)
Constant
396.512307
3.796634
104.437851
0.000001
SOA
-0.053936
0.007638
-7.061800
0.000001
Saliency
3.384938
5.149956
0.657275
0.511039
Bruce & Tsotos (4)
Constant
397.597723
3.696890
107.549231
0.000001
SOA
-0.054002
0.007638
-7.070286
0.000001
Saliency
1.299525
5.140937
0.252780
0.800450
Itti, Koch & Niebur (4)
Constant
396.251191
3.389356
116.910450
0.000001
SOA
-0.053964
0.007636
-7.067030
0.000001
Saliency
4.851707
4.806551
1.009395
0.312842
Table 3.
Sample results of linear regression. Model Zhai & Shah, Method 2, SOA = 75 ms
β
Std. Error
t value
Pr(>|t|)
Constant
415.019703
5.163265
80.379322
0.000000 ***
Saliency
-17.271956
9.029303
-1.912878
0.055965 **
The dependency between reaction time and saliency value
was noticed at a statistically signiicant level only for:
A – Zhai & Shah mode l, SOA = 75 ms, method 2 (p = 0.04) –
Table 3
B – Zhai & Shah mode l,
independently of SOA, method 4
(p = 0.06 – slightly above signiicance threshold) – Table 2
C – SUN DOG, independently of SOA, method 3 (p = 0.03) –
Table 2
For case A (Table 3) the dependency was negative – a shorter
reaction time for higher saliency was noticed. This result is consis
-
tent with the expectations. For cases B (Table 4) and C (Table 2)
the dependency was positive. The results could induce that the
Zhai & Shah and SUN DOG models better simulate the subject’s
behavior. To obtain conclusive results further experiments should
be conducted.
Remarks
Analysis of the obtained results allowed to draw a few conclu-
sions. Firstly, the inluence of SOA on the response time was
observed. This effect, however, was
not consistent with expec
-
tations. It turned out that, regardless of the level of conspicuity
on the locus where the stimulus appeared, for the shorter SOA
subjects responded much more slowly. This results is indepen
-
dent of method used to estimate saliency. There may be several
reasons for such situation but an exact solution would require
further research. Based on information obtained from the subjects,
it can be assumed that most of them for the SOA = 75 ms did not
notice the stimulus appeared on the image. Some respondents
argued that the stimulus appeared together
with the picture and
the response was dificult because there was no licker effect,
visible in the appearance of stimuli in the longer SOA. Perhaps
the time of 75 ms was too short for respondents to prepare the
reaction. Although attention has already made a move in the
direction of the object, the subjects did not respond immediately
T
elema
tics
56
Effectiveness analysis of selected attention models
For method 1, SOA 600ms subjects
obtained shorter reaction
times for regions with higher saliency. It can mean that 600 ms
was enough time to process the picture and attention mechanisms
connected with saliency appeared. Thus, the results were similar
to expected ones.
For method 2, SOA 75 ms, the Zhai & Shah model, the
expected results were obtained. It can show that this model,
although is the simplest one, can make the better prediction.
Maybe such level of complexity is enough to predict human at-
tention movement during free viewing of pictures. On the other
hand, for method 4, Zhah & Shai model and for method 2 SUN
DOG model, in general, results show that the higher level saliency
gets, the longer reaction times are. This is opposite to authors
predictions and opposite to results described previously. The
proposed methods of estimating saliency
and assumptions about
the order of regions in term of level of saliency may be far from
ideal. This issues should be examined in future.
In the rest of the models there was no correlation between
saliency and reaction time. In could be connected with the fact
that the measurement of the reaction times is a complex task
and the results are often noisy. Furthermore, the reaction time
is prone to many distracters:
tiredness, test conditions, time of
the day etc. Therefore, a large amount of samples is required.
The described experiment was performed on only 20 subjects
and 300 images and should be treated as a preliminary study.
It could be assumed that to obtain
more conclusive results the
experiment should be conducted on more subjects and an ex-
tended image database.
The proposed methods of calculating the saliency value in
Do'stlaringiz bilan baham: 
