O. Le Meur and Z. Liu, Saccadic model of eye movements for free-viewing condition, Vision Research, 2015.
We propose a new framework to predict visual scanpaths of observers while they freely watch a visual scene. The visual fixations are inferred from bottom-up saliency and several oculomotor biases. Bottom-up saliency is represented by a saliency map whereas the oculomotor biases (saccade amplitudes and saccade orientations) are modeled using public eye tracking datasets. Our experiments show that the simulated scanpaths exhibit similar trends of human eye movements in a free-viewing condition. The generated scanpaths are more similar to human scanpaths than those generated by two existing methods. In addition, we show that computing saliency maps from simulated visual scanpaths allows to outperform existing saliency models.
Saccadic model of eye movements for free-viewing condition
1. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Saccadic model of eye movements for
free-viewing condition
Olivier Le Meur 1
Zhi Liu 1,2
olemeur@irisa.fr
1
IRISA - University of Rennes 1, France
2
School of Communication and Information Engineering, Shanghai University, China
April 28, 2015
1 / 52
2. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Preamble
The following slides rely heavily upon the following paper:
ª O. Le Meur & Z. Liu, Saccadic model of eye movements for
free-viewing condition, Vision Research, 2015,
doi:10.1016/j.visres.2014.12.026.
Acknowledgments:
This work is supported in part by a Marie Curie International Incoming Fellowship within the 7th European Community Framework
Programme under Grant Nos. 299202 and 911202, and in part by the National Natural Science Foundation of China under Grant
Nos. 61171144 and 61471230.
2 / 52
3. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Outline
1 Visual attention
2 Saccadic model of eye movement
3 Conclusion & Perspective
3 / 52
4. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Visual Attention
1 Visual attention
Presentation
Computational models
Performance
Conclusion
4 / 52
5. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Introduction to visual attention (1/3)
Natural visual scenes are cluttered
and contain many different objects
that cannot all be processed
simultaneously.
Where is Waldo, the young boy
wearing the red-striped shirt...
Amount of information coming
down the optic nerve 108
− 109
bits per second
Far exceeds what the brain is
capable of processing...
5 / 52
6. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Introduction to visual attention (2/3)
WE DO NOT SEE EVERYTHING AROUND US!!!
Visual attention
Posner proposed the following definition (Posner, 1980). Visual atten-
tion is used:
ª to select important areas of our visual field (alerting);
ª to search a target in cluttered scenes (searching).
There are several kinds of visual attention:
ª Overt visual attention: involving eye movements;
ª Covert visual attention: without eye movements (Covert
fixations are not observable).
6 / 52
7. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Introduction to visual attention (3/3)
Bottom-Up vs Top-Down
ª Bottom-Up: some things draw attention reflexively, in a
task-independent way (Involuntary, Very quick, Unconscious);
ª Top-Down: some things draw volitional attention, in a
task-dependent way (Voluntary; Very slow; Conscious).
7 / 52
8. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Introduction to visual attention (3/3)
Bottom-Up vs Top-Down
ª Bottom-Up: some things draw attention reflexively, in a
task-independent way (Involuntary, Very quick, Unconscious);
ª Top-Down: some things draw volitional attention, in a
task-dependent way (Voluntary; Very slow; Conscious).
Computational models of Bottom-up overt visual attention
7 / 52
9. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Computational models of Bottom-up visual
attention (1/3)
Most of the computation models of visual attention have been
motivated by the seminal work of Koch and Ullmann (Koch and
Ullman, 1985).
ª a plausible computational
architecture to predict our
gaze;
ª a set of feature maps are
processed in a massively
parallel manner;
ª a single topographic saliency
map.
8 / 52
10. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Computational models of Bottom-up visual
attention (2/3)
Taxonomy of models:
ª Information Theoretic
models;
ª Cognitive models;
ª Graphical models;
ª Spectral analysis
models;
ª Pattern classification
models;
ª Bayesian models.
Extracted from (Borji and Itti, 2013).
9 / 52
11. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Computational models of Bottom-up visual
attention (3/3)
Cognitive models:
as faithful as possible to
the Human Visual System
(HVS)
ª inspired by cognitive
concepts;
ª based on the HVS
properties.
Extracted from (Borji and Itti, 2013).
10 / 52
12. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Performance on still images (1/3)
The requirement of ground truth
ª Eye tracker:
ª A panel of
observers;
ª An appropriate
protocol.
Adapted from (Judd et al., 2009).
11 / 52
13. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Performance on still images (2/3)
ª Discrete fixation map f i
for the ith
observer:
f i
(x) =
M
k=1
δ(x − xk)
where M is the number of fixations and xk is
the kth
fixation.
ª Continuous saliency map S:
S(x) =
1
N
N
i=1
f i
(x) ∗ Gσ(x)
where N is the number of observers and Gσ is
a 2D Gaussian function.
More details in (Le Meur and Baccino, 2013).
12 / 52
14. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Performance on still images (3/3)
Performance in terms of linear correlation, extracted from (Borji
et al., 2012).
ª more than 30 models.
13 / 52
15. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Conclusion (1/1)
The picture is much clearer than 10 years ago!
BUT...
ª How far are we from human performance?∗
ª How to take advantage of top-down influences?∗
ª What is the best score?∗
ª What is the most representative dataset?∗
Important aspects of our visual system are clearly overlooked;
Current models implicitly assume that eyes are equally likely to
move in any direction;
Systematic biases are not taken into account;
The temporal dimension is not considered (static saliency map)...
∗ Open challenges of visual attention, tutorial CVPR’13, Borji et al.
14 / 52
16. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Visual Attention
2 Saccadic model of eye movement
Presentation
Existing saccadic models
Systematic tendencies
Proposed model
Results
A focus on...
Limitations
15 / 52
17. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Presentation (1/1)
Eye movements are composed of fixations and saccades. A sequence
of fixations is called a visual scanpath.
The fundamental assumption is that scanpaths can be described by a
first-order Markov process, i.e. each eye fixation only depends on the
previous one.
ª The seminal work of (Ellis and Smith, 1985, Stark and Ellis,
1981) described a probabilistic approach where the eye
movements are modelled as a first-order Markov process.
16 / 52
18. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Existing saccadic models (1/6)
ª Saliency models and eye movement
models are combined (Itti and Koch,
2000, Itti et al., 1998).
• winner-take-all algorithm,
• an attended area is inhibited for
approximately 500-900 ms,
• jumps from one salient location to
the next in approximately 30-70
ms.
the model is deterministic, given a
set of of parameters, history of
recent fixation locations...,
the model does not reproduce the
systematic biases of our
oculomotor system.
17 / 52
19. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Existing saccadic models (2/6)
ª (Brockmann and Geisel, 2000)’s
model is related to L´evy flights
which is a particular type of
random walk with a step length
that follows a heavy-tailed
distribution.
• The model is stochastic,
• The model generates
scanpaths similar in their
nature to L´evy flights, i.e.
possess a power law
dependency in their magnitude
distribution..
Saliency information is not
used,
Inhibition is not used.
Top: Model scanpath generated by a
finite variance random walk
(Gaussian system); Bottom: Same
as top, except here the scanpath is a
L´evy flights (Cauchian system).
18 / 52
20. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Existing saccadic models (3/6)
ª (Boccignone and Ferraro, 2004) extended Brockmann’s work,
and modeled eye gaze shifts by using L´evy flights constrained by
the saliency.
• The model is based on a saliency map,
• The jump has a higher probability to occur if the target site is
strongly connected in terms of saliency.
Orientation bias is not used,
Inhibition is not used.
19 / 52
21. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Existing saccadic models (4/6)
ª (Wang et al., 2011)’s model is based on foveal images.
• Visual working memory.
• Distribution of saccade amplitudes is taken into account to jump
from one fixation to another.
• Residual information are used to infer the next fixation location.
Orientation bias is not used,
Time course of IoR is not considered.
20 / 52
22. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Existing saccadic models (5/6)
ª (Tavakoli et al., 2013)
• The model is based on a Markovian process of order one.
• Visited locations are kept in order to inhibit immediate return of
attention to those locations.
• Several parameters are learned from human eye tracking data,
such as p(d) which represents the probability of a jump with
length d. p() is a Gaussian mixture.
Orientation bias is not used,
Time course of IoR is not considered.
21 / 52
23. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Existing saccadic models (6/6)
ª (Liu et al., 2013)
• The model is based on a Markovian process of order one.
• Saliency map.
• L´evy flight to reproduce the distribution of saccade amplitudes.
• Semantic content (transition matrix learned by using Hidden
Markov Model).
Orientation bias is not used,
Time course of IoR is not considered.
Assuming a Markov process and independence between the three
factors:
22 / 52
24. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (1/7)
Saccade amplitudes & Saccade orientations on natural scenes
ª Short saccades around 1
to 3 degrees of visual
angle;
ª Short saccades are more
frequent than long ones.
23 / 52
25. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (2/7)
Saccade amplitudes & Saccade orientations on natural scenes
ª Anisotropic shape;
ª More horizontal saccades
than vertical ones;
ª Very few diagonal
saccades.
23 / 52
26. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (3/7)
Saccade amplitudes & Saccade orientations on natural scenes
ª Joint distribution of
saccade amplitudes and
orientations;
ª Strong horizontal bias
and small saccade.
23 / 52
27. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (4/7)
Horizontal and vertical cross sections of the probability distribution
for horizontal saccades and vertical saccades
ª Saccades in the same
direction as the
precedent one, i.e. with
an angle of 0◦
, are more
likely than saccades in
the opposite direction;
ª Upward vertical saccades
are more likely than
downward saccades
(consistent with (Greene
et al., 2014, Tatler and
Vincent, 2009)).
24 / 52
28. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (5/7)
Saccade amplitudes & Saccade orientations on natural scenes
From top-left to bottom-right: Le Meur, Bruce, Kootstra and Judd’s
fixation datasets. 25 / 52
29. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (6/7)
Center bias on natural scenes
ª Visual fixations are
neither distributed
evenly nor randomly.
26 / 52
30. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Systematic tendencies (7/7)
Saccade amplitudes & Saccade orientations on webpages
ª Joint distribution of
saccade amplitudes and
orientations;
ª Strong horizontal bias in
the rightward direction;
ª F-shaped pattern.
From (Shen and Zhao, 2014)’s eye fixation dataset.
27 / 52
31. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (1/10)
So, what are the key ingredients for successfully designing a saccadic
model?
ª The model has to be stochastic: the subsequent fixation cannot be
completely specified (given a set of data).
ª The model has to generate plausible scanpaths that are similar to
those generated by humans in similar conditions: distribution of
saccade amplitudes and orientations, center bias...
ª Inhibition of return has to be considered: time-course, spatial decay...
ª Fixations should be mainly located on salient areas.
28 / 52
32. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (2/10)
Let I : Ω ⊂ R2
→ R3
an input image and xt a fixation point at time
t.
We consider the 2D discrete conditional probability
p (x|xt−1, · · · , xt−T ) which is composed of three terms:
p (x|xt−1, . . . , xt−T ) ∝ pBU (x)pB(d, φ)pM (x, t) (1)
ª pBU : Ω → [0, 1] is the grayscale saliency map;
ª pB(d, φ) represents the joint probability distribution of saccade
amplitudes and orientations. d is the saccade amplitude between two
fixation points xt and xt−1 (expressed in degree of visual angle), and
φ is the angle (expressed in degree between these two points);
ª pM (x, t) represents the memory state of the location x at time t.
This time-dependent term simulates the inhibition of return.
29 / 52
33. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (3/10)
Bottom-up saliency map
ª pBU is the bottom-up saliency map.
• Computed by GBVS model (Harel et al., 2006). According to
(Borji et al., 2012)’s benchmark, this model is among the best
ones and presents a good trade-off between quality and
complexity.
• pBU (x) is constant over time. (Tatler et al., 2005) indeed
demonstrated that bottom-up influences do not vanish over time.
30 / 52
34. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (4/10)
Joint probability distribution of saccade amplitudes and orientations
ª pB(d, φ) represents the joint probability distribution of saccade
amplitudes and orientations.
• We define di and φi the distance and the angle between each
pair of successive fixations respectively (several eye fixation
datasets were used).
pB(d, φ) =
1
n
n
i=1
Kh(d − di, φ − φi) (2)
where, n is the total number of samples (n = 105727) and Kh is
a two-dimensional Gaussian kernel. h = (hd, hφ) is the kernel
bandwidth.
Separate bandwidths were used for angle and distance
components. The two bandwidth parameters are chosen
optimally based on the linear diffusion method proposed
by (Botev et al., 2010).
31 / 52
35. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (5/10)
Joint probability distribution of saccade amplitudes and orientations
ª pB(d, φ) represents the joint probability distribution of saccade
amplitudes and orientations.
32 / 52
36. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (6/10)
Memory effect and inhibition of return (IoR)
ª pM (x, t) represents the memory effect and inhibition of return
(IoR) of the location x at time t.
• We assume that the IoR effect disappears after T = 8 fixations.
Considering that a fixation duration lasts 300 ms on average, an
attended location could be refixated after 2.4 seconds (Mannan
et al., 1997, Samuel and Kat, 2003);
• We also assume that the spatial IoR effect declines as a
Gaussian function Φσi (d) with the Euclidean distance d from
the attended location (Bennett and Pratt, 2001);
• The temporal decline of the IoR effect is simulated by a simple
linear model.
33 / 52
37. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (7/10)
Memory effect and inhibition of return (IoR)
ª pM (x, t) represents the memory effect and inhibition of return
(IoR) of the location x at time t:
pM (x, t) =
1, t = 0
pM (x, t − 1) − I(y|xt) +
ξ
k=1
R(y|xt−k) otherwise
(3)
where y ∈ Ω represent all the possible locations in the input image. . is
used to clip the value to the range [0, 1]. ξ is the number of visual fixations
to consider: ξ is equal to min(T, t − 1) with T the number of fixations
needed for an area to recover its initial saliency.
I and R represent the inhibition and the recovery functions, respectively:
I(y|z) = Φσi ( y − z ) (4)
R(y|z) =
Φσi ( y − z )
T
(5)
where Φσi (d) is the 2D Gaussian representing the spatial effect of the IoR.
34 / 52
38. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (8/10)
Memory effect and inhibition of return (IoR)
Illustration of the temporal evolution of pM (x, t) for an image defined
over the space Ω = [1...256] × [1...256] and for an unique fixation
point having the spatial coordinate (128, 128). This fixation is
represented by the red cross. We consider, in this example, that the
number of fixation T to recover the original saliency is equal to 4.
35 / 52
39. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (9/10)
Selecting the next fixation point
ª Optimal next fixation point (Bayesian ideal searcher proposed
by (Najemnik and Geisler, 2009)):
x∗
t = arg max
x∈Ω
p (x|xt−1, · · · , xt−T ) (6)
Problem: this approach does not reflect the stochastic behavior
of our visual system and may fail to provide plausible
scanpaths (Najemnik and Geisler, 2008).
ª Rather than selecting the best candidate, we generate Nc = 5
random locations according to the 2D discrete conditional
probability p (x|xt−1, · · · , xt−T ).
The location with the highest saliency gain is chosen as the next
fixation point x∗
t .
36 / 52
40. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Proposed model (10/10)
Selecting the next fixation point
Probability of saccade targeting in retinocentric space by considering that
the previous fixation xt−1 is the centre of the plot. The red crosses are the
Nc candidates that have been randomly drawn according to the shown
probability. From left to right: Nc = {5, 10, 15}.
When Nc = 1, the randomness is maximal, whereas, when Nc increases,
the stochastic behavior of the proposed method is getting less important.
37 / 52
41. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (1/8)
The relevance of the proposed approach is assessed with regard to
the plausibility, the spatial precision of the simulated scanpath
and ability to predict saliency areas.
ª Do the generated scanpaths present the same oculomotor biases
as human scanpaths?
ª What is the similarity degree between predicted and human
scanpaths?
ª Could the predicted scanpaths be used to form relevant saliency
maps?
38 / 52
42. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (2/8)
Are the simulated scanpaths plausible?
ª Protocol:
• We assume that the simulated scanpaths are obtained in a
context of purely free viewing ⇒ top-down effects are not taken
into account.
• For each image in Bruce’s and Judd’s datasets, we generate 20
scanpaths, each composed of 10 fixations ⇒ 224600 generated
visual fixations.
• We assume that the visual fixation duration is constant. So,
considering an average fixation duration of 300ms, 10 fixations
represent a viewing duration of 3s.
• Bottom-up saliency maps are computed by GBVS model (Harel
et al., 2006).
39 / 52
43. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (3/8)
Are the simulated scanpaths plausible?
Top row: Bruce’s dataset. Bottom row: Judd’s dataset.
40 / 52
44. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (4/8)
Are the simulated scanpaths plausible?
Impact of the oculomotor constraints (spatial and orientation),
WTA+IoR
41 / 52
45. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (5/8)
What is the similarity degree between predicted and human scanpaths?
There are a few methods for comparing scanpaths:
string-edit (Privitera and Stark, 2000), Dynamic Time Warp
algorithm (DTW) (Gupta et al., 1996, Jarodzka et al., 2010). More
details in (Le Meur and Baccino, 2013).
ª We use DTW’s method.
ª For a given image, 20 scanpaths each composed of 10 fixations
are generated. The final distance between the predicted
scanpath and human scanpaths is equal to the average of the 20
DTW scores.
The closer to 0 the value DTW , the more similar the scanpaths.
42 / 52
46. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (6/8)
What is the similarity degree between predicted and human scanpaths?
ª Five models are evaluated.
ª The error bars correspond to the SEM (Standard Error of the Mean).
ª DTW = 0 when there is a perfect similarity between scanpaths.
ª There is a significant difference between the performances of the proposed
model and (Boccignone and Ferraro, 2004)’s model (paired t-test,
p << 0.01).
ª As expected, the lowest performances are obtained by the random model.
43 / 52
47. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (7/8)
Scanpath-based saliency map
ª We compute, for each image, 20 scanpaths, each composed of
10 fixations.
ª For each image, we created a saliency map by convolving a
Gaussian function over the fixation locations.
(a) original image; (b) human saliency map; (c) GBVS saliency map; (d)
GBVS-SM saliency maps computed from the simulated scanpaths.
44 / 52
48. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Results (8/8)
Scanpath-based saliency map
45 / 52
49. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Saliency map, randomness and webpages (1/3)
ª Influence of the saliency map:
Top2-SM: we aggregated the saliency maps of GBVS and RARE2012 models
through a simple average. (Le Meur and Liu, 2014) demonstrated that a simple
average of the top 2 saliency maps, computed by GBVS and RARE2012 models,
significantly outperforms the best saliency models.
46 / 52
50. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Saliency map, randomness and webpages (2/3)
ª Randomness:
The maximal randomness is obtained when Nc = 1.
47 / 52
51. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Saliency map, randomness and webpages (3/3)
ª Is the model valid for webpages? By using the eye fixation
dataset of (Shen and Zhao, 2014), we show that oculomotor
biases on webpages differ from those observed on natural scenes:
(a) Distribution of saccade amplitudes; (b) distribution of saccade
orientations; (c) joint distribution of saccade orientations and amplitudes.
ª There is a strong tendency for making horizontal small saccades in the
rightward direction. This tendency is known as the F-bias (Buscher et al.,
2009).
48 / 52
52. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Limitations of the proposed model
Still far from the reality...
ª We do not predict the fixation durations. Some models could be
used for this purpose (Nuthmann et al., 2010, Trukenbrod and
Engbert, 2014).
ª Second-order effect. We assume that the memory effect occurs
only in the fixation location. However, are saccades independent
events? No, see (Tatler and Vincent, 2008).
ª High-level aspects such as the scene context are not included in
our model.
ª Should we recompute the saliency map after every fixations?
Probably yes...
ª Randomness (Nc) should be adapted to the input image. By
default, Nc = 5.
ª Is the time course of IoR relevant? Is the recovery linear?
49 / 52
53. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Conclusion
3 Conclusion & Perspective
50 / 52
54. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Conclusion & Perspective
Improvements
ª Dealing with our model’s limitations.
ª A new metric to evaluate the similarity between scanpaths.
ª How to integrate top-down effects?
ª How to attribute greater meaning to fixations (Bruce, 2014,
Follet et al., 2011, Unema et al., 2005)
Applications
ª How can we use predicted scanpath in computer vision
applications?
ª Is there an application to computational medicine?
51 / 52
55. Saccadic model
O. Le Meur
Visual attention
Presentation
Computational
models
Performance
Conclusion
Saccadic model
of eye movement
Presentation
Existing saccadic
models
Systematic
tendencies
Proposed model
Results
A focus on...
Limitations
Conclusion &
Perspective
Thanks for your attention
52 / 52
56. Saccadic model
O. Le Meur
References
References
P. J. Bennett and J. Pratt. The spatial distribution of inhibition of return:. Psychological Science, 12:76–80, 2001.
G. Boccignone and M. Ferraro. Modelling gaze shift as a constrained random walk. Physica A: Statistical Mechanics and its
Applications, 331:207 – 218, 2004. ISSN 0378-4371. doi: http://dx.doi.org/10.1016/j.physa.2003.09.011.
A. Borji and L. Itti. State-of-the-art in visual attention modeling. IEEE Trans. on Pattern Analysis and Machine Intelligence, 35:
185–207, 2013.
A. Borji, D. N. Sihite, and L. Itti. Quantitative analysis of human-model agreement in visual saliency modeling: A comparative
study. IEEE Transactions on Image Processing, 22(1):55–69, 2012.
Z.I. Botev, J.F. Grotowski, and D. P. Kroese. Kernel density estimation via diffusion. The annals of Statistics, 38(8):2916–2957,
2010.
D. Brockmann and T. Geisel. The ecology of gaze shifts. Neurocomputing, 32(1):643–650, 2000.
Neil D. B. Bruce. Towards fine-grained fixation analysis: Distilling out context dependence. In Proceedings of the Symposium on
Eye Tracking Research and Applications, ETRA ’14, pages 99–102, New York, NY, USA, 2014. ACM. ISBN
978-1-4503-2751-0. doi: 10.1145/2578153.2578167. URL http://doi.acm.org/10.1145/2578153.2578167.
Georg Buscher, Ed Cutrell, and Meredith Ringel Morris. What do you see when you’re surfing? using eye tracking to predict salient
regions of web pages. In Proceedings of CHI 2009. Association for Computing Machinery, Inc., April 2009. URL
http://research.microsoft.com/apps/pubs/default.aspx?id=76826.
S. R. Ellis and J. D. Smith. Patterns of statistical dependency in visual scanning, chapter Eye Movements and Human Information
Processing, pages 221–238. Elsevier Science Publishers BV, (eds) Amsterdam, North Holland Press, 1985.
B. Follet, O. Le Meur, and T. Baccino. New insights into ambient and focal visual fixations using an automatic classification
algorithm. i-Perception, 2(6):592–610, 2011.
Harold H. Greene, James M. Brown, and Barry Dauphin. When do you look where you look? a visual field asymmetry. Vision
Research, 102(0):33 – 40, 2014. ISSN 0042-6989. doi: http://dx.doi.org/10.1016/j.visres.2014.07.012. URL
http://www.sciencedirect.com/science/article/pii/S0042698914001710.
L. Gupta, D. L. Molfese, R. Tammana, and P. G. Simos. Nonlinear alignment and averaging for estimating the evoked potential.
IEEE Transactions on Biomedical Engineering, 43(4):348–356, 1996.
J. Harel, C. Koch, and P. Perona. Graph-based visual saliency. In Proceedings of Neural Information Processing Systems (NIPS),
2006.
52 / 52
57. Saccadic model
O. Le Meur
References
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, May 2000.
L. Itti, C. Koch, and E. Niebur. A model for saliency-based visual attention for rapid scene analysis. IEEE Trans. on PAMI, 20:
1254–1259, 1998.
H. Jarodzka, K. Holmqvist, and K. Nystrom. A vector-based, multidimensional scanpath similarity measure. In ETRA, pages
211–218, 2010.
T. Judd, K. Ehinger, F. Durand, and A. Torralba. Learning to predict where people look. In ICCV, 2009.
C. Koch and S. Ullman. Shifts in selective visual attention: towards the underlying neural circuitry. Human Neurobiology, 4:
219–227, 1985.
O. Le Meur and T. Baccino. Methods for comparing scanpaths and saliency maps: strengths and weaknesses. Behavior Research
Method, 45(1):251–266, 2013.
O. Le Meur and Z. Liu. Saliency aggregation: Does unity make strength? In ACCV, 2014.
H. Liu, D. Xu, Q. Huang, W. Li, M. Xu, and S. Lin. Semantically-based human scanpath estimation with hmms. In ICCV, 2013.
S.K. Mannan, K.H. Ruddock, and D.S. Wooding. Fixation patterns made during brief examination of two-dimensional images.
Perception, 26(8):1059–1072, 1997.
J. Najemnik and W.S. Geisler. Eye movement statistics in humans are consistent with an optimal strategy. Journal of Vision, 8(3):
1–14, 2008.
J. Najemnik and W.S. Geisler. Simple summation rule for optimal fixation selection in visual search. Vision Research, 42:
1286–1294, 2009.
Antje Nuthmann, Tim J. Smith, Ralf Engbert, and John M. Henderson. CRISP: A Computational Model of Fixation Durations in
Scene Viewing. Psychological Review, 117(2):382–405, April 2010. URL
http://www.eric.ed.gov/ERICWebPortal/detail?accno=EJ884784.
M. I. Posner. Orienting of attention. Quarterly Journal of Experimental Psychology, 32:3–25, 1980.
C. M. Privitera and L. W. Stark. Algorithms for defining visual regions-of-interest: Comparison with eye fixations. IEEE
Transactions on Pattern Analysis and Machine Intelligence (PAMI), 22:970–982, 2000.
A.G. Samuel and D. Kat. Inhibition of return: a graphical meta-analysis of its time course and an empirical test of its temporal
and spatial properties. Psychonomic Bulletin & Review, 10(4):897–906, 2003.
Chengyao Shen and Qi Zhao. Webpage saliency. In ECCV. IEEE, 2014.
52 / 52
58. Saccadic model
O. Le Meur
References
L. W. Stark and S. R. Ellis. Scanpaths revisited: cognitive models direct active looking, pages 193–226. Lawrence Erlbaum
Associates, Hillsdale, NJ, 1981.
B.W. Tatler and B. T. Vincent. The prominence of behavioural biases in eye guidance. Visual Cognition, Special Issue: Eye
Guidance in Natural Scenes, 17(6-7):1029–1059, 2009.
B.W. Tatler and B.T. Vincent. Systematic tendencies in scene viewing. Journal of Eye Movement Research, 2:1–18, 2008.
B.W. Tatler, R. J. Baddeley, and I.D. Gilchrist. Visual correlates of fixation selection: effects of scale and time. Vision Research,
45:643–659, 2005.
H.R. Tavakoli, E. Rahtu, and J. Heikkika. Stochastic bottom-up fixation prediction and saccade generation. Image and Vision
Computing, 31:686–693, 2013.
HansA. Trukenbrod and Ralf Engbert. Icat: a computational model for the adaptive control of fixation durations. Psychonomic
Bulletin & Review, 21(4):907–934, 2014. ISSN 1069-9384. doi: 10.3758/s13423-013-0575-0. URL
http://dx.doi.org/10.3758/s13423-013-0575-0.
P.J.A. Unema, S. Pannasch, M. Joos, and B.M. Velichkovsky. Time course of information processing during scene perception: the
relationship between saccade amplitude and fixation duration. Visual Cognition, 12(3):473–494, 2005.
W. Wang, C. Chen, Y. Wang, T. Jiang, F. Fang, and Y. Yao. Simulating human saccadic scanpaths on natural images. In CVPR,
2011.
52 / 52