Image Quality-aware Deep Networks Ensemble
for Efficient Gender Recognition in the Wild
Mohamed Selim
1
, Suraj Sundararajan
1
, Alain Pagani
2
and Didier Stricker
1,2
1
Augmented Vision Lab, Technical University Kaiserslautern, Kaiserslautern, Germany
2
German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany
Keywords:
Gender, Face, Deep Neural Networks, Quality, In the Wild.
Abstract:
Gender recognition is an important task in the field of facial image analysis. Gender can be detected using
different visual cues, for example gait, physical appearance, and most importantly, the face. Deep learning
has been dominating many classification tasks in the past few years. Gender classification is a binary classifi-
cation problem, usually addressed using the facial image. In this work, we present a deep and compact CNN
(GenderCNN) to estimate the gender from a facial image. We also, tackle the illumination and blurriness that
appear in still images and appear more in videos. We use Adaptive Gamma Correction (AGC) to enhance the
contrast and thus, get more details from the facial image. We use AGC as a pre-processing step in gender
classification in still images. In videos, we propose a pipeline that quantifies the blurriness of an image using
a blurriness metric (EMBM), and feeds it to its corresponding GenderCNN that was trained on faces with
similar blurriness. We evaluated our proposed methods on challenging, large, and publicly available datasets,
CelebA, IMDB-WIKI still images datasets and on McGill, and Point and Shoot Challenging (PaSC) videos
datasets. Experiments show that we outperform or in some cases match the state of the art methods.
1 INTRODUCTION
Gender classification is a very important problem in
facial analysis. For humans, the face provides one
of the most important sources for gender classifica-
tion. Besides faces; clothes, physical characteristics
and gait (Ng et al., 2012) can provide information that
identify the gender of a person. While this problem is
a routine task for our brain, it is a challenging task for
computers. Identifying gender from faces has huge
potential in fields like face recognition, biometrics,
advertising and surveillance. Millions of images and
videos are uploaded every day to the Internet. These
images and videos are captured using a variety of de-
vices ranging from mobile phones to DSLR cameras,
under varying conditions. These variations in the cap-
ture process, result in variations in headpose, illumi-
nation, resolution and noise, making gender recogni-
tion from faces challenging (Ng et al., 2012). Gender
recognition on videos is more challenging, due to the
presence of blurriness in the video frames. A gen-
der recognition learning-based method in general in-
volves face detection, feature extraction and finding
the gender from the features (Ng et al., 2012).
In the past few years, deep learning has been dom-
inating different classification tasks (Levi and Hass-
ner, 2015). Convolutional Neural Network (CNN) is
the most widely used neural network for visual recog-
nition systems and natural language processing. Deep
CNNs came into the spotlight in 2012, when a deep
CNN called AlexNet (Krizhevsky et al., 2012) outper-
formed traditional machine learning methods on the
ImageNet Large Scale Visual Recognition Challenge
(ILSVRC) (Russakovsky et al., 2015). ImageNet is
a 1000-class classification challenging competition.
Deeper CNN architectures like VGG (16 or 19 lay-
ers) (Simonyan and Zisserman, 2014) and Residual
Networks (152 layers) (He et al., 2015) keep domi-
nating the challenge. Gender recognition is a binary
classification problem, however it is very challenging
in unconstrained environments, where different varia-
tions exist in the images. It even gets more challeng-
ing in videos due to blurriness or quality compared to
still images.
In this work, we propose a CNN with 3 convolu-
tion layers. Some approaches like (Mansanet et al.,
2016), uses location of the eyes to transform faces
to a canonical pose with eyes located in the same
horizontal line. Our CNN accepts non-aligned faces
as input to predict gender. We evaluate our net-
Selim, M., Sundararajan, S., Pagani, A. and Stricker, D.
Image Quality-aware Deep Networks Ensemble for Efficient Gender Recognition in the Wild.
DOI: 10.5220/0006626103510358
In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 5: VISAPP, pages
351-358
ISBN: 978-989-758-290-5
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
351

Figure 1: Sample gender detection from our proposed ap-
proach using AGC and EMBM GenderCNNs. The video
frame illumination is bad, however AGC improves the fa-
cial image illumination and shows the details of the face.
work on image and video datasets. Due to the chal-
lenges available in data nowadays, new challenging
”in the wild datasets” have been introduced in the
past few years. In our work, the still images datasets
used are IMDB-WIKI (500k images, 2015)(Rothe
et al., 2016) and CelebA (200k images, 2015). Video
datasets used are McGillFaces (35 videos, 10k video
frames, 2015) (Demirkus et al., 2016; Demirkus et al.,
2014; Demirkus et al., 2015) and PaSC (2802 videos,
663846 frames, 2013) (Beveridge et al., 2013). The
meta-data provided in some datasets used face de-
tectors that existed during the time of publishing the
datasets. However, newer face detectors have been
introduced to the community although the problem of
face detection have been studied for many years. The
face detector presented in (Mathias et al., 2014) works
in very challenging situations such as dark or blurred
faces. We decided to use it in our work.
Poorly illuminated faces or blurred ones exist
in still images datasets and exist on a larger scale
in videos datasets. In order to tackle these chal-
lenges, we propose using an adaptive gamma correc-
tion (AGC) (Rahman et al., 2016) method for pre-
processing the faces. We evaluate the impact of us-
ing AGC faces as input to the network. Due to the
movement of the subject or the camera, blurriness
appears in videos. Manually filtering the sharp and
blurry faces is a time consuming process. Hence an
automated method is necessary to separate the faces
based on sharpness. One way of achieving this is to
use a image blurriness metric. We tackle the chal-
lenge of motion blur by using the image blurriness
metric described in (Guan et al., 2015) (EMBM), to
group together faces that are similar in sharpness.
The EMBM values are used to split the pre-processed
faces into groups and separate CNNs are trained for
each group. Each group has faces that fall within a
specific EMBM range.
1.1 Contributions
Our contributions in this work are
• We propose a compact, yet accurate CNN for gen-
der recognition
• We tackle illumination, and quality issues in still
images by pre-processing the faces using AGC
(Rahman et al., 2016).
• For videos, we propose a blurriness-aware Gen-
derCNNs pipeline to detect gender on both sharp
and blurred video frames
We evaluated our approach on challenging, large
and publicly available still images and videos datasets
McGill (Demirkus et al., 2016) , CelebA (Liu et al.,
2015), PaSC (Beveridge et al., 2013) , and IMDB-
Wiki (Rothe et al., 2015) datasets. We outperform
state of the art methods or we perform second-best in
some cases.
2 RELATED WORK
In this section we provide an overview of gender
recognition methods. We start by a brief overview of
face detection methods and pre-processing techniques
which can be used to enhance the input image before
feeding it to a gender classifier. Then, gender recog-
nition methods are discussed.
2.1 Face Detection
Face detection has been widely studied in computer
vision. It is very common that in any application that
involves facial image analysis, it starts with face de-
tection. The Viola and Jones face detector has been
widely used (Viola and Jones, 2001). However, it
has some limitations as it can be limited to frontal
faces, or it can fail in case of occlusions. In 2010,
Deformable Parts Model (DPM) gained popularity in
face detection(Felzenszwalb et al., 2010). In 2014,
Mathias et. al. (Mathias et al., 2014) introduced a
DPM-based face detector. It works well in harsh and
unconstrained environments. We use their face detec-
tor in our approach, as we work with ”in the Wild”
datasets.
2.2 Pre-processing Techniques
Pre-proccessing an image can refer to either image
enhancement or restoration. Image enhancement can
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
352

be required to get the most details available in the im-
age. The details can be lost due to capturing condi-
tions like illumination or the quality of the capturing
device itself. The capturing conditions are not always
optimal for example in images taken by amateurs us-
ing smartphones. Image enhancement is carried out
to improve the image before analyzing it. Enhanc-
ment can be carried out to improve contrast, satu-
ration, or brightness of an image. The illumination
conditions affects the image contrast, consequently,
important details of an image can be lost(Rahman
et al., 2016). Global methods for image enhance-
ment like histogram equalization can result in over or
under-enhanced image, which also results in losing
details of the image. Local methods use neighbouring
pixels to overcome the problems in global methods,
however, they are computationally expensive. Hy-
brid methods use both local and global information
form the image to improve the image, like Adaptive
Histogram Equalization (AHE) or Contrast Limited
Adaptive Histogram Equalization (CLAHE), however
they don’t perform well on various problems, like
dark or bright images.
An Adaptive Gamma Correction (AGC) method
is proposed in (Rahman et al., 2016), and it is able
to enhance dark or bright images. In short, the AGC
method classifies the image first as high or low con-
trast, and then classifies it as dark or bright. AGC
can enhance the images that need enhancement with-
out over or under-enhancing them. For details of the
AGC method please refer to (Rahman et al., 2016).
A comparison of different image enhancement tech-
niques is shown in figure 2. HE over enhanced the
input image. Contrast stretching didn’t improve the
face, it is still dark. CLAHE over enhanced some face
regions, however AGC improved the image without
over or under-enhancing it. Facial details are more
visible. We decided to investigate the use of AGC in
our approach as a pre-processing step.
Figure 2: various contrast enhancing techniques. Over or
under-enhancement in b,c,d. We use AGC (Rahman et al.,
2016) in our approach.
2.3 Gender Recognition
Face has been widely used for gender detection as
shown in the survey (Ng et al., 2012). Earlier works
were constrained to frontal faces. Recent methods
handels varying head pose, illumination, and cap-
turing location. Recently deep learning have been
used in many applications, when the ILSVRC (Rus-
sakovsky et al., 2015) was won by Deep Convolu-
tional Neural Networks (Krizhevsky et al., 2012).
Recent works that use CNNs to solve the prob-
lem of gender classification along with other facial
attributes prediction can be found in (Liu et al., 2015)
and (Ranjan et al., 2016). In (Liu et al., 2015), au-
thors introduced using two known CNNs architec-
tures, LNet and ANet. Both networks were pre-
trained separately and jointly fine tuned. The LNet
was used to localize the face and ANet was used
to extract features for attribute prediction. One at-
tribute was gender. They reported accuracy of 98%
for CelebA and 94% for LFWA datasets.
Later the work in (Ranjan et al., 2016) was in-
troduced. It uses AlexNet initialized with ImageNet
(Russakovsky et al., 2015) weights. They used in-
termediate layers of the network for feature extrac-
tion. The features were classified using SVM. They
were able to reach comparable results with (Liu et al.,
2015) using much less data for training the network.
Another work that uses a deep CNN and SVM for
age and gender recognition, is described in (Levi and
Hassner, 2015). The CNN used has three convolution
layers, 3 fully connected layers and 2 drop layers (for
training). The dataset used for evaluation was Adi-
ence (Eidinger et al., 2014). Adience dataset contains
26,000 images of 2284 subjects. The dataset is con-
strained with limited pose variations and no changes
in location. Approaches with and without oversam-
pling are described in the work. The method using
oversampling achieved a gender recognition accuracy
of 86.8% on Adience.
A deep CNN architecture was proposed in (Si-
monyan and Zisserman, 2014). The study indicates
that the accuracy of image recognition tasks improved
when the depth of the network was increased to upto
19 layers. Two deep architectures called VGG-16 and
VGG-19 were proposed in the work. A face descrip-
tor based on the VGG-16 architecture, was proposed
in (Parkhi et al., 2015), for face recognition. The use
of the descriptor can be investigated in gender recog-
nition.
Gender recognition from still images gathered
more attention than that in videos. In (Wang et al.,
2015), a temporal coherent face desriptor was in-
troduced. Faces from the video frames are used
to create one descriptor, and that descriptor is used
to identify gender using a SVM. The datasets used
in this work were McGillFaces (Demirkus et al.,
2016) and NCKU-driver (Demirkus et al., 2016).
The datasets had variations in illumination, back-
Image Quality-aware Deep Networks Ensemble for Efficient Gender Recognition in the Wild
353

Figure 3: Proposed pipeline for gender detection for still images.
Table 1: Details of the Gender CNN.
Layer Parameters
conv1
num output: 96,
kernel size: 18,
stride: 4
pool1, pool2, pool3
pooling type: MAX,
kernel size: 3,
stride: 2
conv2
num output: 256,
pad: 2,
kernel size: 5,
group: 2
conv3
num output: 512,
pad: 2,
kernel size: 3,
group: 2
fc1 num output: 6000
fc2 num output: 2
norm1, norm2
(LRN layers)
local size: 5,
alpha: 0.0001,
beta: 0.75
ground, head pose, facial movements and expres-
sions. The method showcased an accuracy of 94.12%
and 96.67%for McGillFaces and NCKU respectively.
3 DEEP LEARNING FOR
GENDER RECOGNITION
3.1 Proposed Network
Deep and well known networks like AlexNet, LeNet,
and others were designed for 1000-class classification
problem in ImageNet (Large Scale Visual Recogni-
tion Challenge (Russakovsky et al., 2015)). The prob-
lem in hand is a binary classification problem, where
it is required to know from the facial image, the gen-
der of the subject. It is not a trivial task, however we
work in the context of faces only. We propose to use
a compact network, that consists of only 3 convolu-
tion layers followed by two fully connected layers. In
the evaluation section, we show that we do not need
a network with many convolution layers to recognize
gender. We show the details of our proposed network
in table 1.
The network takes facial images of size 200 x 200
pixels. The first convolution layer conv1 has 96 fil-
ters of size 18 x 18 pixels. The second convolution
layer conv2 has 256 filters of size 5 x 5 pixels. The
last convolution layer conv3 has 512 filters of size 3
x 3 pixels. The first fully connected layer has size
of 6000, followed by the second layer fc2 of size 2.
It represents male and females, and is followed by a
SoftMax layer to predict the gender of the input im-
age.
3.2 Gender Recognition in Still Images
In order to perform gender classification on still im-
ages, we use our proposed network model with Adap-
tive Gamma Correction as a pre-processing step. In
figure 3, we show an overview of the pipeline. First,
we detect the face in the input image using the face
detector in (Mathias et al., 2014). The detected face
is cropped, and pre-processed using AGC (Rahman
et al., 2016). The pre-processed face is fed to the net-
work and the gender is predicted. Using AGC helps
in enhancing the image, in case that is required. If the
image is already good enough, the AGC will not dis-
tort it with over or under-enhancing. More challenges
exist in video frames and we discuss gender detection
in video frames in the next subsection.
3.3 Gender Recognition in Videos
Videos captured using mobile phones, low resolution
cameras or Point and Shoot cameras are challenging
for gender recognition, due to the presence of vary-
ing illumination and motion blur. These challenges
are not usual in still images. The presence of poorly
illuminated frames, results in dark faces being fed to
the machine learning algorithm. We tackle the illu-
mination problem using the AGC, as we do in the still
images pipeline. Motion blur is a common problem in
videos captured by in an unstable manner, for exam-
ple, handheld cameras, or where the subject is mov-
ing fast and the capturing device is not fast enough.
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
354

Figure 4: Gender detection for video frames, considering detected face EMBM value.
Figure 5: Effect of AGC on dark faces. a Original im-
ages and their EMBM (Guan et al., 2015) values. b AGC-
enhanced (Rahman et al., 2016) images and their EMBM
values. Faces extracted from PaSC(Beveridge et al., 2013).
We can clearly see these effects in the PaSC dataset,
especially the handheld videos. In order estimate gen-
der in bad quality or in blurred video frames, we use
a blurriness metric to quantify the amount of mo-
tion blur. We use EMBM metric proposed in (Guan
et al., 2015), to separate faces into groups accord-
ing to the EMBM value. However, in case of dark
faces, the EMBM value is mistakenly estimated to be
very blurry, which is not true. The EMBM value can
be also wrong in case of dark face with bright back-
ground, as the EMBM value estimates the sharpeness
between the face and the background. We apply the
AGC to enhance the facial image, and then evaluate
EMBM. This gives a more accurate and reliable value
of the EMBM as the details of the face are restored
by AGC. Example images of evaluating EMBM with
and without applying AGC are shown in figure 5. We
can see in the figure that the AGC-enhanced images
gives more reliable EMBM values.
After computing the EMBM value, we split the
images into groups, and for each group, we use our
proposed network model to estimate the gender for
the face under its corresponding blurriness group.
Based on our experimental results, we split the im-
ages into two groups based on EMBM threshold. The
threshold value can differ based on the data used.
4 EVALUATION
In this section, we discuss the evaluation of our pro-
posed method for gender classification. The next two
subsections discuss the datasets used in our evalua-
tion, and shows the results on these datasets. In our
results we compare to state of the art methods, and
we show that we either outperform other methods or
we reach comparably close accuracy. Our experiment
setup is as follows, we use 80% of the dataset for
training and 20% for testing. We make sure that there
is no overlap between the training and testing data. In
order to have a fair comparison with other learning-
based state of the art methods, we consider fine tuning
their models similarly as we do for our model. How-
ever, our CNN model is trained from scratch.
4.1 Still Images
In 2015, Liu et. al. (Liu et al., 2015) introduced the
CelebA dataset. The dataset consists of 202,599 im-
ages of 10,177 celebrities collected from the internet
along with 40 facial attributes for each image. One of
the attributes is gender.
Another still images dataset we work with is the
IMDB-Wiki(Rothe et al., 2015) dataset. It was also
introduced in 2015. It consists of 460,723 images
from IMDB, and 62,328 from Wikipedia. The dataset
is mainly intended for apparent age estimation, how-
ever, it also has the gender ground truth information.
Sample Images from the datasets are shown in figure
7. We consider the images from IMDb and Wikipedia
as separate datasets in our evaluations.
Experimental results are shown in table 3. We
evaluated our proposed method on the following still
images datasets, IMDB-Wiki (Rothe et al., 2015) and
CelebA (Liu et al., 2015). We compare our results to
Image Quality-aware Deep Networks Ensemble for Efficient Gender Recognition in the Wild
355

Table 2: Results and comparisons on still images datasets. Best result is shown in Bold, second best is shown in Bold Italic.
IMDB Wiki CelebA
Age & Gender CNN (Levi and Hassner, 2015) 87.2% 94.44% 97.27%
VGG face desc (Simonyan and Zisserman, 2014) & SVM 57.05% 81.37% 91.99%
LNet+ANet (Liu et al., 2015) - - 98%
GenderCNN wo AGC (Ours) 87.34% 94.75% 97.35%
GenderCNN w AGC (Ours) 87.46% 94.76% 97.38%
(a) CelebA (Liu et al., 2015)
(b) Wiki (Rothe et al., 2015)
(c) IMDB (Rothe et al., 2015)
Figure 6: Still images datasets samples.
other state of the art methods. We outperform other
methods on the IMDB-Wiki dataset. On the CelebA
dataset, the work in (Liu et al., 2015) still has the best
accuracy value of 98%, however we score the second
best accuracy in the table of value 97.38%. Another
important point is the use of nearly 30k images less
for training. In our work, we use a network that has 3
convolution layers, however, in (Liu et al., 2015) they
use two deep CNNs to perform facial attribute predic-
tions. We get the score of LNet and ANet (Liu et al.,
2015) method from their paper (Liu et al., 2015). Pre-
processing the images with AGC shows slight im-
provement in still images datasets. The next subsec-
tion discusses the videos datasets used and our eval-
uation results on them. The temporal coherent face
descriptor cannot be applied on still images datasets,
as it works only with videos.
4.2 Videos
The McGill (Demirkus et al., 2016; Demirkus et al.,
2014; Demirkus et al., 2015) dataset, created in 2015,
has a total of 60 videos of 31 female and 29 male sub-
jects. Videos have 300 frames each with a resolution
of 640 x 480, and varies in terms of location (indoor
or outdoor), illumination conditions, head pose, oc-
clusions and background clutter. Only a subset of 35
videos (10308 frames) have been shared with us. Fig-
ure 7(b) shows sample video frames from two sepa-
rate videos.
The Point and Shoot Face Recognition Challenge
(PaSC) (Beveridge et al., 2013) dataset, created in
2013, includes still images and videos. We perform
our evaluations only on the videos from this dataset.
The dataset has a total of 2802 videos of 265 different
people shot at six different locations. There are equal
number of control videos and handheld videos. The
control videos were shot using a high quality camera
at a resolution of 1920x1080 pixels (Full HD) with
a tripod. The handheld videos were shot using vari-
ous devices with resolution ranging from 640x480 to
1280x720. The dataset did not contain gender infor-
mation needed for our evaluations. We were able to
add the gender information manually. Of the 265 peo-
ple, 143 were male and 122 were female. The varying
locations (indoor and outdoor), resolution, illumina-
tions, head poses, makes the PaSC (Beveridge et al.,
2013) dataset ideal for evaluating gender recognition.
The evaluation results are shown in table 3. We
evaluated the GenderCNN with and without AGC,
along with AGC and EMBM GenderCNNs on the
McGill and PaSC datasets. On the two datasets, we
outperform the Temporal coherent face desriptor used
by (Wang et al., 2015) for gender detection by a big
margin. On the McGill dataset, using VGG face de-
scriptor and a SVM (Simonyan and Zisserman, 2014)
gives the best accuracy, however, we score the second
best accuracy value. On the PaSC dataset, we out-
perform the state of the art methods we compared to,
our proposed methods give the best and second best
scores. Using GenderCNN with AGC pre-processing
gave the best results on the control subset of the PaSC
dataset. The experimental results show that on the
challenging subset, the handheld, using the blurriness
metric EMBM to train different CNNs gave the best
result. Figure 1 shows gender detection on a sample
frame from a handheld video with bright background.
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
356

Table 3: Results and comparisons on videos datasets. Best result is shown in Bold, second best is shown in Bold Italic.
McGill PaSC Control PaSC Handheld
Age & Gender CNN (Levi and Hassner, 2015) 84.4% 92.41% 92.62%
VGG face desc (Simonyan and Zisserman, 2014) & SVM 92.13% 90.45% 91.26%
Temporal gender detector (Wang et al., 2015) 71.42% 86.45% 88.04%
GenderCNN wo AGC (Ours) 90.14% 92.1% 92.84%
GenderCNN w AGC (Ours) 90.4% 93.6% 94%
AGC +EMBM GenderCNNs (Ours) - 93.31% 94.2%
(a) McGill (Demirkus et al., 2016)
(a) PaSC (Beveridge et al., 2013)
Figure 7: Videos datasets sample frames.
The effect of AGC in improving the contrast of the
face is visible.
5 CONCLUSION
In this paper a deep and compact CNN, with only 3
convolution layers was proposed for gender recogni-
tion from non aligned faces. We evaluated our Gen-
derCNN on recent and challenging images and videos
datasets. The still images used in our evaluation were
IMDB-Wiki(Rothe et al., 2015) dataset and CelebA
(Liu et al., 2015). We outperform the state of the art
or match their accuracy values with small difference.
In order to have a fair comparison, we trained the
learning-based state of the art methods using the dif-
ferent datasets. We proposed using Adaptive Gamma
Correction, AGC (Rahman et al., 2016) in our gen-
der classification pipeline as a pre-processing step. It
improved the results on still images datasets.
On the videos datasets, we proposed a different
pipeline that tackles challenges found in videos. One
main challenge was the image quality degradation
due to blurriness. We propose quantifying the blur-
riness using EMBM (Guan et al., 2015) and training
different CNNs based on this metric. We observed
that poorly illuminated faces in video frames give
misleading EMBM values, we tackled that using the
pre-processing using AGC. Our proposed pipeline for
gender detection in videos showed improvement on
the challenging handheld videos of the PaSC dataset.
The work presented in this paper opens the door for
future ideas to investigate methods to tackle the blur-
riness found in the ”in the wild” videos. One possi-
ble idea would be investigating the usefulness of de-
blurring techniques and their impact on gender classi-
fication of facial images analysis.
ACKNOWLEDGMENT
This work has been partially funded by the Univer-
sity project Zentrums f
¨
ur Nutzfahrzeugtechnologie
(ZNT).
REFERENCES
Beveridge, J. R., Phillips, P. J., Bolme, D. S., Draper, B. A.,
Givens, G. H., Lui, Y. M., Teli, M. N., Zhang, H.,
Scruggs, W. T., Bowyer, K. W., Flynn, P. J., and
Cheng, S. (2013). The challenge of face recognition
from digital point-and-shoot cameras. In 2013 IEEE
Sixth International Conference on Biometrics: The-
ory, Applications and Systems (BTAS), pages 1–8.
Demirkus, M., Precup, D., Clark, J. J., and Arbel, T.
(2014). Probabilistic Temporal Head Pose Estimation
Image Quality-aware Deep Networks Ensemble for Efficient Gender Recognition in the Wild
357

Using a Hierarchical Graphical Model, pages 328–
344. Springer International Publishing, Cham.
Demirkus, M., Precup, D., Clark, J. J., and Arbel, T. (2015).
Hierarchical temporal graphical model for head pose
estimation and subsequent attribute classification in
real-world videos. Computer Vision and Image Un-
derstanding, 136:128 – 145. Generative Models in
Computer Vision and Medical Imaging.
Demirkus, M., Precup, D., Clark, J. J., and Arbel, T. (2016).
Hierarchical spatio-temporal probabilistic graphical
model with multiple feature fusion for binary facial at-
tribute classification in real-world face videos. IEEE
Transactions on Pattern Analysis & Machine Intelli-
gence, 38(6):1185–1203.
Eidinger, E., Enbar, R., and Hassner, T. (2014). Age and
gender estimation of unfiltered faces. Trans. Info. For.
Sec., 9(12):2170–2179.
Felzenszwalb, P. F., Girshick, R. B., McAllester, D., and
Ramanan, D. (2010). Object detection with discrimi-
natively trained part-based models. IEEE Trans. Pat-
tern Anal. Mach. Intell., 32(9):1627–1645.
Guan, J., Zhang, W., Gu, J. J., and Ren, H. (2015). No-
reference blur assessment based on edge modeling.
J. Visual Communication and Image Representation,
29:1–7.
He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep
residual learning for image recognition. CoRR,
abs/1512.03385.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-
agenet classification with deep convolutional neural
networks. In Advances in neural information process-
ing systems, pages 1097–1105.
Levi, G. and Hassner, T. (2015). Age and gender classifi-
cation using convolutional neural networks. In IEEE
Conf. on Computer Vision and Pattern Recognition
(CVPR) workshops.
Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learn-
ing face attributes in the wild. In Proceedings of In-
ternational Conference on Computer Vision (ICCV).
Mansanet, J., Albiol, A., and Paredes, R. (2016). Local
deep neural networks for gender recognition. Pattern
Recogn. Lett., 70(C):80–86.
Mathias, M., Benenson, R., Pedersoli, M., and Van Gool, L.
(2014). Face detection without bells and whistles. In
ECCV.
Ng, C. B., Tay, Y. H., and Goi, B.-M. (2012). Vision-
based human gender recognition: A survey. CoRR,
abs/1204.1611.
Parkhi, O. M., Vedaldi, A., and Zisserman, A. (2015). Deep
face recognition. In British Machine Vision Confer-
ence.
Rahman, S., Rahman, M. M., Abdullah-Al-Wadud, M., Al-
Quaderi, G. D., and Shoyaib, M. (2016). An adaptive
gamma correction for image enhancement. EURASIP
Journal on Image and Video Processing, 2016(1):35.
Ranjan, R., Patel, V. M., and Chellappa, R. (2016). Hyper-
face: A deep multi-task learning framework for face
detection, landmark localization, pose estimation, and
gender recognition. CoRR, abs/1603.01249.
Rothe, R., Timofte, R., and Gool, L. V. (2015). Dex: Deep
expectation of apparent age from a single image. In
IEEE International Conference on Computer Vision
Workshops (ICCVW).
Rothe, R., Timofte, R., and Gool, L. V. (2016). Deep ex-
pectation of real and apparent age from a single im-
age without facial landmarks. International Journal
of Computer Vision (IJCV).
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,
S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,
Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).
ImageNet Large Scale Visual Recognition Challenge.
International Journal of Computer Vision (IJCV),
115(3):211–252.
Simonyan, K. and Zisserman, A. (2014). Very deep con-
volutional networks for large-scale image recognition.
CoRR, abs/1409.1556.
Viola, P. and Jones, M. (2001). Rapid object detection us-
ing a boosted cascade of simple features. In Proceed-
ings of the 2001 IEEE Computer Society Conference
on Computer Vision and Pattern Recognition. CVPR
2001, pages 511–518.
Wang, W. C., Hsu, R. Y., Huang, C. R., and Syu, L. Y.
(2015). Video gender recognition using temporal co-
herent face descriptor. In 2015 IEEE/ACIS 16th Inter-
national Conference on Software Engineering, Artifi-
cial Intelligence, Networking and Parallel/Distributed
Computing (SNPD), pages 1–6.
VISAPP 2018 - International Conference on Computer Vision Theory and Applications
358