Gender Classification using the Gaze Distributions of Observers
on Privacy-protected Training Images
Michiko Inoue, Masashi Nishiyama and Yoshio Iwai
Graduate School of Engineering, Tottori University, Tottori, Japan
Keywords:
Gender Classification, Training Images, Gaze Distributions, Privacy-protection.
Abstract:
We propose a method for classifying the gender of pedestrians using a classifier trained by images containing
privacy-protection of the head region. Recently, manipulated training images containing pedestrians have
been required to protect the privacy of pedestrians. In particular, the head regions of the training images are
manipulated. However, the accuracy of gender classification decreases when privacy-protected training images
are directly used. To overcome this issue, we aim to use the human visual ability to correctly discriminate
males from females even though the head regions have been manipulated. We measure the gaze distributions
of observers who view pedestrian images and use them to pre-process gender classifiers. The experimental
results show that our method using gaze distribution improved the accuracy of gender classification when the
head regions of the training images have been manipulated with masking, pixelization, and blur for privacy-
protection.
1 INTRODUCTION
Video surveillance cameras are now installed in var-
ious public spaces such as airports, stations, and
shopping malls. Gender classification using pedes-
trian images acquired from video surveillance cam-
eras is becoming widespread. The gender classifi-
cation of pedestrian images enables the distribution
of pedestrian gender to be collected for the develop-
ment of various applications such as crime prevention
and product marketing. The existing methods (Su-
dowe et al., 2015; Schumann and Stiefelhagen, 2017)
achieve high accuracy using deep learning techniques
for pedestrian attribute classification. However, in the
existing methods, the collection of a large number
of training images is required to achieve good deep
learning performance.
When collecting training images, we must care-
fully handle the privacy of the human subjects in the
images. Training images include personal informa-
tion that makes human subjects feel nervous about
allowing their images to be acquired. In particular,
we must pay attention to how their faces are handled.
Furthermore, these training images will be used re-
peatedly over a long period of time to improve the ac-
curacy of gender classification. Therefore, techniques
for protecting the privacy of the subjects in the im-
ages are required when collecting training images for
gender classification.
To protect the privacy of subjects in pedestrian im-
ages, the head regions are generally manipulated. For
example, head regions in magazine advertisements
are masked, head regions in television interviews are
pixelized, and head regions in Google Street View are
blurred. Furthermore, methods to perform more com-
plex manipulations for privacy protection have been
proposed (Joon et al., 2016; Zhang et al., 2014; Ya-
mada et al., 2013; Ribaric et al., 2016; Oh et al.,
2017). However, existing methods do not fully dis-
cuss the privacy protection of training images in gen-
der classification. When they are directly applied to
training images, we believe that the accuracy of gen-
der classification will be substantially decreased. In
fact, (Ruchaud et al., 2015) showed that this decrease
happens when the head regions of training images are
manipulated with masks, pixelization, and blur to pro-
tect the privacy of the subjects.
In this paper, we focus on the human visual abil-
ity to distinguish gender to improve the accuracy
of gender classification using privacy-protected train-
ing images. The existing method (Nishiyama et al.,
2018) revealed that observers look at important fea-
tures when distinguishing the gender of pedestrians
in images. In that approach, gaze distributions were
measured when observers viewed the pedestrian im-
ages. The existing method assigned weights to the
Inoue, M., Nishiyama, M. and Iwai, Y.
Gender Classification using the Gaze Distributions of Observers on Privacy-protected Training Images.
DOI: 10.5220/0008876101490156
In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages
149-156
ISBN: 978-989-758-402-2; ISSN: 2184-4321
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
149

MaleFemale
Figure 1: Observers can correctly determine gender when
the head regions of pedestrian images are completely
masked to protect the privacy of the subjects.
training images as preprocessing before feature ex-
traction so that a classifier could be trained to em-
phasize the head region, which is where the gaze of
the observers gathered. However, when the existing
method is directly applied to privacy-protected train-
ing images, the accuracy of gender classification un-
fortunately decreases. The reason for this issue is that
the head regions have been manipulated to protect the
privacy of the pedestrians in the training images, even
though head regions increase the accuracy of gender
classification. We need to consider how to overcome
this problem caused by privacy-protected training im-
ages. Here, we similarly focus on human visual abili-
ties. When the head regions of pedestrian images are
completely masked, observers can still correctly dis-
tinguish gender in many cases. We believe that the
observers look at other important features obtained
from regions other than the head, such as the shape
and appearance of the torso, in the pedestrian image.
As illustrated in Figure 1, observers can correctly dis-
criminate gender; for example, it is a woman when the
torso of the subject is rounded, and it is a man when
the torso of the subject is muscular.
In this paper, we propose a method to improve the
accuracy of gender classification when the head re-
gions in the training images are manipulated for pri-
vacy protection. We make the following two contri-
butions:
• We reveal the important regions where the gaze
of observers tends to gather in privacy-protected
images. We measure the gaze distributions of the
observers when the head regions of the subjects
are masked.
• We confirm whether or not the accuracy of gender
classification is improved using these gaze distri-
butions. We use the gaze distributions in feature
extraction for gender classifiers.
To compare the important regions of the masked im-
ages, we also perform experiments in which the head
regions are not masked.
2 RELATED WORK
2.1 Privacy Protection
As described in (Joon et al., 2016; Flammini et al.,
2013; Campisi, 2013), the privacy protection of
subjects contained in images is an important issue,
and has attracted attention in recent years. Various
privacy-protection methods have been proposed for
all stages in the process of recognition, such as dur-
ing pedestrian image acquisition or when classifica-
tion is performed using the pedestrian image. To pre-
vent recognition in conventional face detection algo-
rithms, a method that requires the user to wear protec-
tive glasses was proposed in (Yamada et al., 2013). To
automatically shield the heads of subjects in images,
a method that incorporates a special mechanism em-
bedded in the camera system was proposed in (Zhang
et al., 2014). The existing methods have the advan-
tage that privacy-protected images are only recorded
in video surveillance systems. However, it is neces-
sary to prepare special equipment for image acquisi-
tion. Existing methods are not suitable for collect-
ing training images. To protect the privacy of faces, a
method that replaces the face of a subject with the face
of a fictitious person has been proposed in (Ribaric
et al., 2016). Although this method opens up new ap-
proaches to privacy-protection, it is not well accepted
by the public. To hide personal features so that they
are not visually discernible in an image for a classifi-
cation task, a method that embeds the original feature
in the image of a fictional person was proposed in (Oh
et al., 2017). This is effective if the algorithm of the
classification task is not changed in the future. How-
ever, the embedded features are not restored when a
newly developed algorithm is applied. Instead, we
consider a pre-processing method for feature extrac-
tion that can enable various classification algorithms
to use privacy-protected training images.
2.2 Human Visual Abilities
In the fields of computer vision and pattern recog-
nition, the use of human visual abilities has widely
progressed. The estimation accuracy of a saliency
map was improved using the distribution of gaze lo-
cations in (Xu et al., 2015). Action recognition and
gaze-attention estimation were performed simultane-
ously using a wearable camera in (Fathi et al., 2012).
Preference estimation was performed using gaze lo-
cations and image features in (Sugano et al., 2014),
an eye-movement pattern for product recommenda-
tion was estimated in (Zhao et al., 2016), and an ob-
ject recognition task was performed using only gaze
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
150

(b)
(a)
S
1
S
2
Figure 2: Examples of stimulus images for measuring gaze
distributions.
distributions in (Karessli et al., 2017). The attributes
of face images were classified using gaze-attention re-
gions in (Murrugarra-Llerena and Kovashka, 2017),
and the attribute classification of fashion clothes im-
ages was performed using a deep learning technique
with gaze distributions in (Sattar et al., 2017). Al-
though the existing methods handle various applica-
tions using gaze, they do not address the application
of privacy protection. Thus, privacy protection in gen-
der classification is a new application for gaze distri-
butions.
3 GAZE DISTRIBUTION OF
OBSERVERS WHEN THE HEAD
REGIONS ARE MASKED
3.1 Stimulus Images
We investigated which regions gather the gaze of ob-
servers when they determine the gender of subjects in
images. Sixteen participants (10 males and 6 females,
average age 22.4±1.0 years) participated in the study.
The head regions and their surrounding regions were
completely masked so that the participants could not
observe the head region of the subject in the images.
Note that we refer to the pedestrian images observed
by the participants as the stimulus images.
We used the following two conditions for the
pedestrian images in the experiments:
S
1
: the head region was masked;
S
2
: the head region was not masked.
We measured S
2
in addition to S
1
for comparison.
Figures 2(a) and (b) show the examples of the stimu-
(a)
S
1
(b)
S
2
Figure 3: Average images of pedestrian images with and
without masked head regions.
65 cm
100 cm
Display
70 cm
Eye tracker
Participant
110–120 cm
Figure 4: Experimental setting for measuring gaze distribu-
tions.
lus images of S
1
and S
2
. We used the CUHK dataset,
which is included in the PETA dataset (Yubin et al.,
2014) as the stimulus images. The size of all stimulus
images was 80 × 160 pixels for both S
1
and S
2
.
The pedestrian regions in the images of the CUHK
dataset were manually aligned. We checked the align-
ment using the average image computed from all
pedestrian images in the CUHK dataset. Figures 3(a)
and (b) show the average images of S
1
and S
2
. In
Figure 3(b) , the black circle that appears at the top
corresponds to the head region. The black ellipse that
appears near the center of the image corresponds to
the torso region. The light gray part that appears at
the bottom of the image corresponds to the foot re-
gion. In Figure 3(a), we see that the head region is
completely masked because the black circle at the top
in Figure 3(b) is not observed.
To control the experimental conditions, the num-
ber of male and female subjects included in the stimu-
lus images was equal. The proportions of all body ori-
entations of the subjects in the stimulus images (front,
back, left and right) were equal. In addition, the same
person did not appear more than once in the stimulus
images. Finally, the number of the stimulus images
was 32 in S
1
and 32 in S
2
.
Figure 4 shows the experimental setting for mea-
suring gaze distributions. The participant was seated
at a position 65 cm from the display. Each partici-
Gender Classification using the Gaze Distributions of Observers on Privacy-protected Training Images
151

(a)
S
1
(b)
S
2
Figure 5: Examples of S
1
and S
2
stimulus images presented
on the display.
pant adjusted the chair height while keeping the eye
height between 110 and 120 cm. The display size
was 24 inches (1, 920 × 1, 080 pixels). We used the
GP3 gaze measurement device (gazepoint), which has
a sampling rate of 60 fps. The specifications of the
device state that its angular resolution is between 0.5
and 1.0 degrees. We enlarged the stimulus image to
480 × 960 pixels. To avoid center bias, we presented
the stimulus images at random positions on the dis-
play. Figures 5(a) and (b) show examples of the S
1
and S
2
stimulus images on the display. The pixel val-
ues of the masked regions in the stimulus images were
set to be the same as those of the display background.
We presented the stimulus images of S
1
for 8 partici-
pants (4 males and 4 females) randomly selected from
all the participants and the stimulus images of S
2
for
the remaining 8 participants (6 males and 2 females).
3.2 Protocol
The procedure for measuring the gaze distribution of
the participants was as follows.
P
1
: We randomly selected a participant.
P
2
: We set the condition for measuring the gaze dis-
tribution to either S
1
or S
2
.
P
3
: We explained how to perform the task of deter-
mining the gender of the subject using an exam-
ple image.
P
4
: We presented a gray image for 2 s.
P
5
: We presented a randomly selected stimulus im-
age for 2 s.
P
6
: We presented a black image for 3 s and asked
the participant to state the gender of the subject
in the stimulus image.
P
7
: We repeated steps P
4
to P
6
until all the stimulus
images were presented.
P
8
: We repeated steps P
1
to P
7
until all the partici-
pants completed the experiment.
Here, we explain our method for generating a gaze
map. We integrated the gaze distributions measured in
P
5
using the existing method (Nishiyama et al., 2018)
for a single gaze map. We used only the gaze loca-
tions of the participants, which the gaze measurement
g
S
1
g
S
2
(a) (b)
Figure 6: Gaze maps of S
1
and S
2
.
device output as fixations. We summed the locations
of the fixations for S
1
or S
2
from all the participants
and all stimulus images. The size of the gaze map was
resized to the size of the stimulus image.
3.3 Analysis of the Measured Gaze
Distributions
Figure 6 shows the gaze maps of S
1
and S
2
. We show
the gaze map g
i
for each i ∈ {S
1
, S
2
}. In the map,
darker regions of the gaze map indicate that the gaze
of the participants were more frequently measured at
these locations than in the lighter regions. Compar-
ing the gaze map g
S
1
of Figure 6(a) with the average
image of Figure 3(a), we see that the gaze locations
of the participants gathered in the torso regions of the
subjects when the head regions were masked. The
body shape, clothes, and bag were observed in the
subjects of the stimulus images. Next, we compared
g
S
2
of Figure 6(b) with the average image of Fig-
ure 3(b). We see that the participants mainly viewed
the head regions of the subjects in the stimulus im-
ages when they were not masked. This tendency is the
same as that of the existing study (Nishiyama et al.,
2018). Note that the gaze locations of the participants
did not gather near the feet regions of the subjects in
both S
1
and S
2
stimulus images.
We describe the accuracy of gender classification
performed by the participants. We counted correct an-
swers when the responses of the participants matched
the gender labels of the stimulus image. The accu-
racy of the participants was 87.0 ± 5.5% for S
1
and
95.9 ± 3.1% for S
2
.
We investigated the differences in the gaze maps
with respect to the participants’ gender. We also in-
vestigated the differences in the stimulus images with
respect to the subjects’ gender. Figure 7(a) shows the
gaze maps generated by each gender of the partici-
pants, and Figure 7(b) shows the gaze maps generated
for each gender of the subjects. In the maps in Fig-
ure 7(a), there is no significant difference between the
genders of the participants in terms of the person re-
gion gaze locations. In Figure 7(b), there was also no
significant difference between the gender of the sub-
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
152

(a) Participants
(b) Subjects
Male
Female
S
1
S
2
Male
Female
Male
Female
Male
Female
S
1
S
2
Figure 7: Gaze maps measured of (a) male and female par-
ticipants and (b) male and female subjects.
Table 1: Correlation coefficients of the gaze maps for males
and females with respect to participants or subjects.
Participants Subjects
S
1
S
2
S
1
S
2
0.88 0.97 0.97 0.92
jects in the maps. For both S
1
and S
2
, we computed
the correlation coefficients between the gaze maps
with respect to participant gender and subject gender.
We used Pearson’s product–moment correlation co-
efficient. Table 1 shows the correlation coefficients
of the gender differences in S
1
and S
2
. We confirmed
that the correlation coefficients between genders were
very high under all combinations of conditions.
4 GENDER CLASSIFICATION
USING THE GAZE
DISTRIBUTIONS FOR
PRIVACY-PROTECTED
TRAINING IMAGES
4.1 Experimental Conditions
We investigated whether or not the accuracy of gen-
der classification was improved using the gaze dis-
tributions for the training images in which the head
regions of the subjects were masked. We used images
from the CUHK dataset as training and test images.
(a)
(b)
T
1
S
T
2
S
Figure 8: Examples of training images with and without
masked head regions.
The stimulus images for gaze measurement were not
included in the training and test images. To evalu-
ate the accuracy of gender classification, we used 10-
fold cross-validation for a total of 2,540 images of the
CUHK dataset. In each cross-validation, the number
of training images was 2,286 (1,143 male images and
1,143 female images), and the number of test images
was 254 (127 male images and 127 female images).
We define the training image conditions as follows:
• T
S
1
: the head regions of the training images were
masked;
• T
S
2
: the head regions of the training images were
not masked.
Figure 8(a) shows examples of the training images of
T
S
1
and Figure 8(b) shows those of T
S
2
.
To confirm the effectiveness of the gaze map for
privacy-protected training images, we compared the
accuracy of gender classification under the following
conditions:
G1 : our method using the masked gaze map g
S
1
;
G2 : the existing method using the un-masked gaze
map g
S
2
;
G3 : a baseline method without the use of the gaze
map.
The number of combinations of conditions of G1, G2,
and G3 for T
S
1
and T
S
2
was six. We avoided gaze
measurement for each training image and each test
image and used only the gaze maps measured from
the stimulus images in Section 3. We used the clas-
sifier with gaze-map based pre-processing described
in (Nishiyama et al., 2018). The pre-processing is
briefly explained below. Large weights are given to
the pixels of the training images at which the gaze of
the participants gathered. In contrast, small weights
Gender Classification using the Gaze Distributions of Observers on Privacy-protected Training Images
153

：G1 ：G2 ：G3
45
70
65
60
55
50
(b) SVM
45
70
65
60
55
50
Accuracy (%)
45
70
65
60
55
50
(a) CNN
(c) LR
Accuracy (%)
Accuracy (%)
Figure 9: Accuracy of gender classification using the
masked training images of T
S
1
.
are given to the pixels at which gaze was not gath-
ered. Specifically, let the pixel value of the gaze map
be g
i
(x, y), where (x , y) is the location of the image.
The range of g
i
(x, y) is [0, 1]. In addition, the pixel
value of the training image is represented as t(x, y).
The pixel value t
0
(x, y) after pre-processing is com-
puted as follows:
t
0
(x, y) = c(g
i
(x, y))t(x, y), (1)
where c( ) is a correction function. We set c(z) =
z
a
+ b. When a > 1, the weight of the gaze map is
emphasized. When a < 1, the weight of the gaze map
is low. Variable b is an offset. We converted the RGB
color space to the HSV color space and weighted only
the V values using the correction function.
After assigning the weights to the training images,
we employed the following classifiers:
• CNN: convolutional neural network. A mini-
CNN (Grigory et al., 2015) with two convolution
layers and two pooling layers was used.
• SVM: linear support vector machine (Corinna
and Vladimir, 1995). The penalty parameter was
set as 1.
60
80
75
65
70
85
60
80
75
65
70
85
60
80
75
65
70
85
(b) SVM
Accuracy (%)
(a) CNN
(c) LR
Accuracy (%)
Accuracy (%)
：G1 ：G2 ：G3
Figure 10: Accuracy of gender classification using the un-
masked training images of T
S
2
.
• LR: logistic regression classifier (Cox, 1958).
The normalization parameter was set as 1.
We did not mask the head regions of the test images.
We weighted the test images using the gaze map us-
ing the same procedure as that used for the training
images.
4.2 Performance of Gender
Classification using the Masked
Training Images
Figure 9 shows the accuracy of gender classification
for the masked training images of T
S
1
. In Figure 9(a),
we see that the accuracy of G1 was slightly better than
that of G3, and the accuracy of G2 was significantly
worse than those of G1 and G3. In Figure 9(b), we
see the same tendencies shown in Figure 9(a). In Fig-
ure 9(c), we see that the accuracy of G1 is superior to
those of G2 and G3, and the accuracy of G2 is slightly
better than that of G3. These results confirm that gaze
map g
S
1
is more effective than g
S
2
when the head re-
gions of the training images are masked.
Figure 10 shows the accuracy of gender classifica-
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
154

(b)
(c)
(a)
T
1
p
T
1
b
T
2
p
T
3
p
T
4
p
T
2
b
T
3
b
T
4
b
Figure 11: Examples of the training images after manip-
ulating the head regions using pixelization and blur while
changing their parameters.
tion for the unmasked training images of T
S
2
. In Fig-
ure 10(a), there is no significant difference in accu-
racy between G1 and G3. We see that the accuracy of
G2 is better than those of G1 and G3. In Figure 10(b),
we see that the accuracy of G1 is worse than those
of G2 and G3 and the accuracy of G2 is better than
that of G3. In Figure 10(c), we see the same tenden-
cies shown in Figure 10(b). These results confirm that
gaze map g
S
2
is more effective than g
S
1
when the head
regions of the training images are not masked. We be-
lieve that it is necessary to switch between the gaze
maps suitable for each condition because the effec-
tiveness of the gaze distributions depends on whether
or not the training images are masked.
4.3 Gender Classification using the
Training Images with Pixelization
and Blur
We investigated the accuracy of gender classification
when the classifiers are trained by images modified
with pixelization and blur for privacy protection. Fig-
ure 11(a) shows an example of a training image with-
out privacy protection, Figure 11(b) shows images
with pixelization, and Figure 11(c) shows images with
Accuracy (%)
(a) Pixelization
(b) Blur
60
80
75
65
70
60
80
75
65
70
Accuracy (%)
T
1
p
T
1
b
T
2
p
T
3
p
T
4
p
T
2
b
T
3
b
T
4
b
：G1 ：G2 ：G3
Figure 12: Accuracy of gender classification using the train-
ing images manipulated with pixelization and blur.
blur. We used four manipulation levels k ∈ {1, 2, 3, 4}
for pixelization T
p
k
and blur T
b
k
. We varied the level
of pixelization (using 16 × 32, 12 × 24, 8 × 16, and
4 × 8 blocks) and the level of Gaussian blur (σ =
3, 7, 11, 15). We used a CNN classifier and the con-
ditions G1, G2, and G3 for the gaze maps described
in Section 4.1.
Figure 12(a) shows the accuracy of gender clas-
sification using the training images with pixelization.
In T
p
1
, T
p
2
, and T
p
3
, G2, G3, and G1 have the highest,
middle, and lowest accuracy, respectively. In contrast,
in T
p
4
, this order is G1, G3, and G2. These results
hence confirm that our method based on gaze map g
S
1
improves the accuracy for training images with a high
level of pixelization.
Figure 12(b) shows the accuracy of gender classi-
fication using training images with blur. In T
p
1
, T
p
2
,
T
p
3
, and T
p
4
, G1, G2, and G3 have the highest, mid-
dle, and lowest accuracy, respectively. We hence con-
firm that our method based on gaze map g
S
1
improves
the accuracy for training images with various levels
of blur.
Gender Classification using the Gaze Distributions of Observers on Privacy-protected Training Images
155

5 CONCLUSIONS
We proposed a method for improving the accuracy of
gender classification using the gaze distribution of hu-
man observers on training images in which the pri-
vacy of the subjects was protected. We used stimu-
lus images with masked head regions and measured
the gaze distributions of observers. We confirmed
that the participants mainly observed the torso re-
gions of the subjects in the stimulus images. Next,
we conducted gender classification experiments us-
ing privacy-protected training images with masking,
pixelization, and blur. The experimental results con-
firm that our method, which uses the gaze map with
masked head regions, improved the accuracy of gen-
der classification. In future work, we intend to con-
tinue developing the method to increase the accuracy
by combining gaze maps with and without masking.
We will expand this investigation into gaze maps with
privacy protection for various classification tasks re-
lated to attributes other than gender. This work was
partially supported by JSPS KAKENHI under grant
number JP17K00238 and MIC SCOPE under grant
number 172308003.
REFERENCES
Campisi, P. (2013). Security and privacy in biometrics: To-
wards a holistic approach. Security and Privacy in
Biometrics, Springer, pages 1–23.
Corinna, C. and Vladimir, V. (1995). Support-vector net-
works. Machine Learning, 20(3):273–297.
Cox, D. R. (1958). The regression analysis of binary se-
quences. Journal of the Royal Statistical Society. Se-
ries B (Methodological), 20(2):215–242.
Fathi, A., Li, Y., and Rehg, J. (2012). Learning to recognize
daily actions using gaze. In Proceedings of 12th Euro-
pean Conference on Computer Vision, pages 314–327.
Flammini, F., Setola, R., and Franceschetti, G. (2013). Ef-
fective surveillance for homeland security: Balancing
technology and social issues. CRC Press.
Grigory, A., Sid-Ahmed, B., Natacha, R., and Jean-Luc, D.
(2015). Learned vs. hand-crafted features for pedes-
trian gender recognition. In Proceedings of 23rd ACM
International Conference on Multimedia, pages 1263–
1266.
Joon, O. S., Rodrigo, B., Mario, F., and Bernt, S. (2016).
Faceless person recognition: Privacy implications in
social media. In Proceedings of European Conference
on Computer Vision, pages 19–35.
Karessli, N., Akata, Z., Schiele, B., and Bulling, A. (2017).
Gaze embeddings for zero-shot image classification.
In Proceedings of IEEE conference on computer vi-
sion and pattern recognition, pages 4525–4534.
Murrugarra-Llerena, N. and Kovashka, A. (2017). Learning
attributes from human gaze. In Proceedings of IEEE
Winter Conference on Applications of Computer Vi-
sion, pages 510–519.
Nishiyama, M., Matsumoto, R., Yoshimura, H., and Iwai,
Y. (2018). Extracting discriminative features using
task-oriented gaze maps measured from observers for
personal attribute classification. Pattern Recognition
Letters, 112:241 – 248.
Oh, S. J., Fritz, M., and Schiele, B. (2017). Adversarial
image perturbation for privacy protection a game the-
ory perspective. In Proceedings of IEEE International
Conference on Computer Vision, pages 1491–1500.
Ribaric, S., Ariyaeeinia, A., and Pavesic, N. (2016). De-
identification for privacy protection in multimedia
content. Image Communication, 47(C):131–151.
Ruchaud, N., Antipov, G., Korshunov, P., Dugelay, J. L.,
Ebrahimi, T., and Berrani, S. A. (2015). The impact
of privacy protection filters on gender recognition. Ap-
plications of Digital Image Processing XXXVIII, page
959906.
Sattar, H., Bulling, A., and Fritz, M. (2017). Predicting
the category and attributes of visual search targets us-
ing deep gaze pooling. In Proceedings of IEEE Inter-
national Conference on Computer Vision Workshops,
pages 2740–2748.
Schumann, A. and Stiefelhagen, R. (2017). Per-
son re-identification by deep learning attribute-
complementary information. In Proceedings of IEEE
Conference on Computer Vision and Pattern Recogni-
tion Workshops, pages 1435–1443.
Sudowe, P., Spitzer, H., and LeibeSudowe, B. (2015). Per-
son attribute recognition with a jointly-trained holis-
tic cnn model. In Proceedings of IEEE International
Conference on Computer Vision Workshop, pages
329–337.
Sugano, Y., Ozaki, Y., Kasai, H., Ogaki, K., and Sato,
Y. (2014). Image preference estimation with a
data-driven approach: A comparative study between
gaze and image features. Eye Movement Research,
7(3):862–875.
Xu, M., Ren, Y., and Wang, Z. (2015). Learning to pre-
dict saliency on face images. In Proceedings of IEEE
International Conference on Computer Vision, pages
3907–3915.
Yamada, T., Gohshi, S., and Echizen, I. (2013). Privacy vi-
sor: Method for preventing face image detection by
using differences in human and device sensitivity. In
Proceedings of International Conference on Commu-
nications and Multimedia Security, pages 152–161.
Yubin, D., Ping, L., Change, L. C., and Xiaoou, T. (2014).
Pedestrian attribute recognition at far distance. In Pro-
ceedings of 22nd ACM International Conference on
Multimedia, pages 789–792.
Zhang, Y., Lu, Y., Nagahara, H., and Taniguchi, R. (2014).
Anonymous camera for privacy protection. In Pro-
ceedings of 22nd International Conference on Pattern
Recognition, pages 4170–4175.
Zhao, Q., Chang, S., Harper, F. M., and J. A. Konstan, J.
(2016). Gaze prediction for recommender systems.
In Proceedings of 10th ACM Conference on Recom-
mender Systems, pages 131–138.
VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications
156