Live Stream Oriented Age and Gender Estimation using Boosted LBP
Histograms Comparisons
Lionel Prevost
1
, Philippe Phothisane
2
and Erwan Bigorgne
2
1
LAMIA, University of the French West Indies and Guiana, Campus de Fouillole, BP 250, 97157 Pointe
`
a Pitre, France
2
Eikeo, 11 rue L
´
eon Jouhaux, 75010 Paris, France
Keywords:
Face Analysis, Boosting, Gender Estimation, Age Estimation.
Abstract:
Research has recently focused on human age and gender estimation because they are useful cues in many ap-
plications such as human-machine interaction, soft biometrics or demographic statistics for marketing. Even
though human perception of other people’s age is often biased, attaining this kind of precision with an auto-
matic estimator is still a difficult challenge. In this paper, we propose a real time face tracking framework that
includes a sequential estimation of people’s gender then age. A single gender estimator and several gender-
specific age estimators are trained using a boosting scheme and their decisions are combined to output a gender
and an age in years. We choose to train all these estimators using local binary patterns histograms extracted
from still facial images. The whole process is thoroughly tested on state-of art databases and video sets. Re-
sults on the popular FG-NET database show results comparable to human perception (overall 70% correct
responses within 5 years tolerance and almost 90% within 10 years tolerance). The age and gender estima-
tors can output decisions at 21 frames per second. Combined with the face tracker, they provide real-time
estimations of age and gender.
1 INTRODUCTION
Humans can glean a wide variety of information from
a face image, including identity, age, gender, and eth-
nicity. Despite the broad exploration of person iden-
tification from face images, there is only a limited
amount of research on how to automatically and ac-
curately estimate demographic information contained
in face images such as age or gender.
Gender identification is a cognitive process
learned and consolidated throughout childhood. It
finally becomes mature in teenage years(Wild et al.,
2000). Children can make good guesses but also use
many social stereotypes such as facial features, hair
type, clothes or interests. Cropped faces without these
external features are generally enough for adult oper-
ators to classify properly men and women. The goal
of an automatic gender estimator is to match adult
human accuracy on cropped face image. As social
stereotypes are too variable, they are not considered
by most existing methods.
Age identification is also learned throughout life
experiences. It is easier to guess someone’s age when
his(her) age range and ethnicity are seen frequently
(Anastasi and Rhodes, 2005) . One’s appearance age
may be altered by his(her) individual growth pattern,
general health, ethnicity, gender, etc. All these param-
eters should be considered to determine someone’s
age with accuracy. But most of the time, facial im-
ages provide enough information about a subject to
be able to estimate his(her) age range.
Automatic age and gender estimators can be
used in many different applications such as human-
computer interaction, security control, demographic
segmentation for marketing studies, etc. Research
teams report good performances on databases well
spread in the FAR (Facial Analysis and Recognition)
community, such as FERET or LFW(Huang et al.,
2007).
In this article, we introduce our own method for
age and gender estimation. They are implemented in
a real-time 3D face tracker derived from the one de-
tailed in (Phothisane et al., 2011) and provide age and
gender estimation (figure 1). The main contributions
of the paper are:
• The estimation of gender using one bi-class
boosted classifier based on multi-block local bi-
nary pattern (MBLBP) histogram comparisons.
• The estimation of age age using several gender
790
Prevost L., Phothisane P. and Bigorgne E..
Live Stream Oriented Age and Gender Estimation using Boosted LBP Histograms Comparisons.
DOI: 10.5220/0004927207900798
In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 790-798
ISBN: 978-989-758-018-5
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
specific age classifiers combined with a weighted
sum rule to output an age .
• The comparison of human perception vs. auto-
matic estimation of age on two databases.
• The study of our estimators in videos using 3D
face tracking. Such experiments on real data ap-
pear rarely in most publications focused on age
and gender estimation.
In section 2, previous methods are introduced, provid-
ing state of the art performances on age and gender
estimation. Section 3 describes the feature extraction
process. Section 4 details boosting training and deci-
sion process for gender and age estimation. Section
5 presents various experiments on common databases
and video sequences to validate our approach. Com-
parisons with state of the art is provided too. Finally,
section 6 concludes and adds some prospects.
2 PREVIOUS WORK
2.1 Gender Recognition
Most published methods use face cues for gender
recognition. The first attempts of automatic gen-
der estimation started in the early 90’s with the
SEXNET(Golomb et al., 1991). This method used a
two-layer neural network trained to classify 30 × 30
facial images. Tests were done on 90 images (45
males, 45 females) and obtained a 8.68% error rate.
At the same time, Cottrell and Metclafe(Cottrell and
Metclafe, 1990) used 160 64 × 64 images (10 males,
10 females). Images were reduced to 40 components
vectors and used to train a single layer neural net-
work. This experiment provided a perfect recogni-
tion rate on the training database. Recently, other
methods have also used gait cues to gather more in-
formation on targeted subjects(Li et al., 2008; Shan
et al., 2008). As our human interaction applications
are aimed at being used at close range, only faces are
visible. The focus here is set on methods using only
face images. Many recent papers report results on the
FERET database. Moghaddam and Yang (Moghad-
dam and Yang, 2002), achieve an overall 3.38% error
rate using a support vector machine with a Gaussian
kernel on low resolution images. Baluja and Row-
ley(Baluja and Rowley, 2007) report comparable re-
sults using simply pixel comparisons in 20 × 20 face
images. This feature extraction process is very inter-
esting because it is not time consuming.
Other studies published results on more uncon-
strained databases, as image sets downloaded from
the web. Shan’s gender estimator(Shan, 2012), ap-
plies SVMs to LBP histograms on 7,443 images of
the LFW database, obtaining a 5.19% error rate.
Shakhnarovich et al.(Shakhnarovich et al., 2002),
Gao and Ai(Gao and Ai, 2009), Kumar et al.(Kumar
et al., 2008) experimented on non publicly available
databases. (Shakhnarovich et al., 2002) uses an ad-
aBoost on Haar filters outputs applied to 30 × 30 im-
ages. Kumar et al.(Kumar et al., 2008) obtain an
8.62% error rate on a 1,954 images database (1,087
males, 867 females) with SVM comparable to those
seen in (Moghaddam and Yang, 2002). Most stud-
ies focus on still image databases using k-fold cross-
validation, and few provide cross-database results.
Makinen and Raisamo (Makinen and Raisamo, 2008)
provide a deep comparison of some state-of-art clas-
sifiers (Neural Network, SVM and various AdaBoost)
on gender estimation. These classifiers are trained on
the FERET database and evaluated on a homemade
”internet” database. Authors show that the mean ac-
curacy is good (80% to 90%) and does not vary sig-
nificantly from one classifier to another. Overall,
only a few experiments were conducted on video se-
quences(Hadid and M., 2008).
2.2 Age Estimation
Estimating an age means to automatically assign an
age to the current subject, whether in years or as an
age interval. It is the reverse action of age modeling
(Ramanathan and Chellappa, 2006). There appear to
be several definitions of “age“ described in (Fu et al.,
2010).
• The actual age is the real age of an individual.
• The perceived age is gauged by another person.
• The appearance age, given by the person’s image.
• The estimated age is given by a computer.
Age estimation can be seen as two different prob-
lems. The first is a regression problem where the esti-
mator has to predict someone’s age as closely as pos-
sible with a year precision. The second aims at clas-
sifying a face image into one of several bins. As an
example, Gao and Ai(Gao and Ai, 2009) use a linear
discriminant analysis on Gabor wavelets and classify
images into 4 bins (baby, child, adult, old). Recently,
Guo et al.(Guo et al., 2009b) studied both questions
using bio-inspired features (BIF) and achieved a 4.77
years accuracy on the FG-NET database. Thukral et
al. (Thukral et al., 2012) report a mean absolute error
(MAE) of 6.2 years on the whole FG-NET database
using geometric features and relevance vector ma-
chines. Luu et al. report a MAE of 4.37 years on
LiveStreamOrientedAgeandGenderEstimationusingBoostedLBPHistogramsComparisons
791
Figure 1: Gender then age estimation process flowchart. The apparent blocks are the 64 (8 rows, 8 columns) regions of
interest. The signature is the 256 × 59 2-uniform LBP normalized histograms matrix.
a subset of FG-NET using active appearance mod-
els and support vector machine regressors (Luu et al.,
2009) and 4.12 years on the complete set with a con-
tourlet appearance model (Luu et al., 2011). Others
report results on the whole database, using methods
such as RUN (Regressor on Uncertain Nonnegative
labels(Yan et al., 2007)) and (Lanitis et al., 2004).
(Guo et al., 2009a) reports a 4.69 years accuracy on
the non-publicly available Yamaha Gender and Age
(YGA) database. As we can see, many studies re-
port results on the FG-NET database which is pub-
licly available (http://www.fgnet.rsunit.com). Most
of them use a ”Leave-One-Subject-Out” evaluation
scheme and their MAE varies between 4 and 6 years.
We decided to perform LBP histogram bin com-
parisons instead of simplistic pixel comparison and to
use an AdaBoost scheme to perform sequential gen-
der then age comparison. We study our method’s be-
havior and compare it to state of art methods on stan-
dard databases.
3 MULTI-SCALE BLOCK LOCAL
BINARY PATTERN
HISTOGRAMS
First, we perform 2D face detection and 3D face
alignment using a real real-time face tracker and pose
estimator described in (Phothisane et al., 2011). Its
precision allows us to track facial features accurately.
Eyes coordinates are used to extract a cropped face
from the source image and normalize it to 128 × 96.
Pose estimation also provides valuable information
and allows to reject images of face too far from a
frontal pose. This rejection is only used on our live
video tests. For still image databases, no rejection is
applied. Then, we compute Multi-scale Block LBP
(MBLBP) and histograms fo MBLBP. The following
subsections describe this process thoroughly.
3.1 Uniform Local Binary Patterns
LBP are commonly used local texture descriptors
(Ojala et al., 2002). Their evolutions include
Multi-resolution Histograms of Local Variation Pat-
terns(Zhang et al., 2005) and Multi-scale Block
LBP(Liao et al., 2007)(MBLBP) which inspired our
method.
The original (scale-1) LBP operator labels the
pixels of an image by thresholding the 3 × 3-
neighborhood of each pixel with the center value p
0
and considering the result as a binary string. For each
p
i
, i = {1, ..., 8} surrounding p
0
in a circular fashion,
the boolean b
i
= 1 if p
i
> p
0
and b
i
= 0 if p
i
6 p
0
is
associated. Using the 8 bits, LBP
p
0
has 256 possible
values. The histogram of the labels can be used as a
texture descriptor.
LBP
p
0
=
∑
i
2
i−1
× b
i
(1)
In MB-LBP, the comparison operator between sin-
gle pixels in LBP is simply replaced with comparison
between average gray-values of sub-regions. Each
sub-region is a square block containing neighboring
pixels (or just one pixel particularly). The whole filter
is composed of 9 blocks. We take the size k of the fil-
ter as a parameter, and k × k denoting the scale of the
MB-LBP operator. For instance, a scale-3 LBP cen-
tered on pixel p
0
uses all the pixels in the 9 ×9 region
surrounding it. The reference region, r
0
is a 3×3 area
around p
0
. Each other r
i
, i = {1, ..., 8} is another 3×3
region encircling r
0
defined as the sum of its pixel val-
ues p. Thus, the b
i
are defined according to the sum
of the pixel values p inside the r
i
regions:
b
i
=
1 if
∑
p∈r
i
p >
∑
p∈r
0
p
0 if
∑
p∈r
i
p 6
∑
p∈r
0
p
(2)
In our method, scale-1, 3, 5 and 9 LBP are com-
puted before conversion into 2-uniform LBP. The k-
uniform LBP are a subset of the original LBP. The
criterion used is the number of circular bit transitions:
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
792
a k-uniform LBP has k or less transitions. For instance
11001111 and 00000001 have both 2 transitions and
thus are 2-uniform LBP. According to (Shan, 2012)
and (Ojala et al., 2002), 2-uniform LBP provide the
majority of seen patterns. There are 58 possible val-
ues of 2-uniform LBP, the remaining values are all
set as non-uniform. We use a look-up table which di-
rectly transforms LBP values into 58+1 different val-
ues. We obtain in the end four 128 × 96 2-uniform
LBP maps for each input image (one for each scale).
3.2 Block Histograms
In order to add spatial information, in a similar fash-
ion of (Shan, 2012) and (Liao et al., 2007), we com-
pute histograms of 2-uniform LBP values contained
in subwindows of 26 × 20 pixels over the different
layers.The subwindows are regularly distributed into
8 rows and 8 columns. In the end we obtain 256 59-
bin histograms. Each 59-bin histogram is normalized
to obtain a unit vector. These 59 × 256 matrix signa-
tures are computed on each face image.
4 BOOSTING AND DECISION
The signatures are used to classify age and gender.
Before any learning, the face database (described
in section 5.1) is labeled, with actual gender and
a perceived age. Each image is mirrored to avoid
asymmetrical bias during the learning process. The
database thus doubles in size. The weak classifiers
f (c, j
1
, j
2
), c = {1, ..., 59}, j
n
= {1, ..., 256}, j
1
6= j
2
are simple comparisons of histogram components
across subwindows. For instance, the c
th
histogram
bin value h
c, j
1
of subwindow j
1
is compared to ev-
ery other c
th
histogram bin value h
c, j
2
, of subwindow
j
2
, j
2
6= j
1
. There are C =
(256×255)
2
×59 = 1, 925, 760
weak classifiers. The C weak classifiers of each train-
ing image are used to build our gender and age esti-
mators.
• if h
c, j
1
> h
c, j
2
, f (c, j
1
, j
2
) = 1
• if h
c, j
1
6 h
c, j
2
, f (c, j
1
, j
2
) = 0
4.1 Gender Estimation
Gender identification is a bi-class segmentation and
age estimation is a multi-class segmentation. For the
gender estimator, a single strong classifier is built
upon the selected weak classifiers using AdaBoost
training scheme (Freund and Schapire, 1996). The
gender strong classifier S
g
combines weak classifiers
outputs using the following equation.
S
g
=
∑
K
k=1
log(
1−e
k
e
k
) × f (c
k
, j
1,k
, j
2,k
)
∑
K
k=1
log(
1−e
k
e
k
)
(3)
K being the number of iterations computed by ad-
aBoost, and e
k
representing the weighted error on the
training database after the k
th
iteration. f (c
k
, j
1,k
, j
2,k
)
is the output of the k
th
selected weak classifier. Then,
the decision is taken by thresholding S
g
.
• if S
g
> 0.5 + δ, subject is a male.
• if S
g
< 0.5 − δ, subject is a female.
With δ being the decision threshold. S
g
values
within the [0.5 − δ, 0.5 + δ] interval are considered
neutral. In our real-time implementation, we set δ =
0.01 and obtain good qualitative results in live demon-
strations.
4.2 Age Estimation
According to the results provided by (Guo et al.,
2009a), age estimators provide better results after be-
ing trained on specific genders. This result is intu-
itive as male and female facial features are not altered
by age in the same way. So, males and females are
segregated in the training databases in order to build
two gender specific age estimators. We build a com-
plete age estimator by training several strong classi-
fiers S
a
, a ∈ {10, 15, ..., 50, 55} which output a real
value like the previous gender strong classifier. Each
S
a
is constructed using specific image selections and
labels. The S
a
learn to classify face images into two
classes: those younger than a, and those older than a.
The output of S
a
is computed by using equation (3).
then, all the strong classifiers outputs are used to com-
pute an over-the-ages score S
age
. The age decision is
made by finding the maximum value and associated
age k ∈ {0, 1, ..., 59, 60} of S
age
(k):
argmax
k
S
age
(k) =
∑
a∈A
(S
a
− .5) ×
1
1 + e
k−a
(4)
The sigmoid fonction in equation (4) is used to nor-
malize the strong classifier outputs and to build a con-
tinuous over-the-ages score. Though simplistic, this
decision is effective.
4.3 Real-time Video Analysis
On still images databases used for validation, the age
and gender estimations are made without any specific
threshold. However, for sequential databases and the
live application, the distribution of gender and age
estimations associated with each target is important.
LiveStreamOrientedAgeandGenderEstimationusingBoostedLBPHistogramsComparisons
793
Both age and gender estimators are implemented in
our real-time facial analysis system, which provides
accurate head pose estimation. The gender estimation
is triggered when the face alignment is considered
satisfactory enough, according to specific regression
score thresholds on each facial landmark. According
to the gender estimator’s decision, the male or the fe-
male age estimator makes the age estimation. To use
the information provided by the face tracking, com-
puted age and gender estimations are collected over
the sequence. These estimations are recorded for each
tracked target, building two unidimensional votes dis-
tributions. Our observations made us choose to model
these distributions with gaussian mixture models, and
to use the E-M algorithm to take decisions. After fit-
ting the age model and the gender model, the most
weighted gaussian bell means are selected as the final
decision.
For the video analysis, every frame is consid-
ered. The computation times of our C++ implementa-
tion were measured on Intel Core i7-2600 hardware:
42+/-1 ms for a MBLBP matrix and 1.5+/-0.1 ms for
12,000 weak classifiers (age and gender). Added up,
these single-threaded processes can be computed at
21 frames per second with one core while the other
cores are dedicated to other tasks such as face track-
ing. The computational load can be lowered by reduc-
ing the MBLBP signature generation frequency, as
two consecutive frames are likely to be only slightly
different.
5 EXPERIMENTS
Our age and gender estimators are compared to other
state of the art methods on still images databases
used in the facial analysis community. The FG-NET
database is used to test our method for age estimation,
and LFW and FERET are used for gender recognition.
Other experiments are conducted on video sequences.
These databases are described in the next subsection.
5.1 Databases
Gender. Labeled Faces in the Wild (LFW) and
FERET are commonly used database among the facial
analysis community, particularly for face and gender
recognition. We randomly select a subset of LFW
to keep a balanced repartition of males and females.
This final selection contains a total of 2,758 images.
The FERET image selection contains the 1,696 im-
ages extracted from the f a and f b subsets. We pro-
vide results on these databases using 4-fold cross val-
idation.
Age. The FG-NET database is an age estimation spe-
cific database. It contains 1002 images of 82 different
people, with age labels going from 0 to 69 years. The
”100” video (”from 0 to 100 years in 150 seconds”) is
available on youtube. During this sequence, 101 peo-
ple from 0 to 100 years old tell their age while fac-
ing the camera. The first frame corresponding to each
subject was extracted and labeled accordingly. The
”0 to 100” database is used to measure age estimation
errors from human operators and automatic system.
Age and Gender. We built our own age and gen-
der face database by using images collected from the
web. The objective was to gather faces with a wide va-
riety of pose, illumination and expression for a large
number of people from various origins. It contains for
now 5,814 images, including 3,366 males and 2,448
females. Ten human operators labeled these images
with the age they perceived. In order to measure ac-
curacy of human age perception, all these operators
participated in a dedicated experiment described in
section 5.2. A test-only database was also collected
in order to have a constant validation set, as the size
of our training dataset aims to grow in size. Even
though this method is inherently subject to bias, the
error margin is measured in two experiments.
The recorded video dataset uses 16 videos of
8 people, including 6 males and 2 females. The
face alignment was considered good enough in 2,086
frames. Every subject was asked to look at the cam-
era then look at specific items in order to capture a
wide range of face poses. The relative low number of
sequences is compensated by the quantity of images.
In order to investigate the estimators’s behaviour to-
wards asymmetric facial appearances, each subject
was captured in two different illumination conditions:
one in ambient lighting and the other with a supple-
mental lateral light source. The experimental results
are provided in the next subsection.
5.2 Results on Still Images
For the following experiments, gender is estimated
first and according to the estimator’s decision, the age
estimation uses either the male or the female features
selection.
Gender Estimation. Experiments on gender estima-
tion were done on the LFW and FERET databases.
Four-fold cross validation tests were proceeded on
both dababases separately, obtaining 90.7% accuracy
on our subset of LFW and 93.4% of correct answers
on the FERET database. The experiments described
in 5.3 measure the gender estimator’s performance in
video sequences.
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
794
Human Age Perception Experiments. Measuring
human errors of age perception would help appreciate
automatic age estimation results. Two distinct mea-
surements are conducted. The first is done on a subset
of the FG-NET database, and the other on the 0 to 100
dataset. Ten people participated in each experiment.
The objective is to measure the accuracy of human
age perception and compare it to current automatic
methods, including ours.
The first experiment uses an age-uniform selection
of 60 clear FG-NET pictures of males and females
from 0 to 69 years old. Human age perception errors
are shown in figure 2. The MAE is 4.9 years with a
standard deviation of 4.6 years. More than 65% of
the errors are below 5 years and almost 90% of the
errors are below 10 years. These values are close to
the performance of the best age estimators published
recently and not far from those reported by (Han et al.,
2013) on the whole FG-NET set (MAE=4.7 years).
For the second experiment, either all images (from
0 to 100 years old) or only a subset (from 0 to 60 years
old) are considered. The human performance is sim-
ilar to the one in the previous test using the FG-NET
subset (figure 3). This provides information about
what kind of accuracy is achievable and reveals that
the performance of the best age estimators are actu-
ally close to human perception.
Age Estimation. To perform fair comparison with
existing methods, a leave-one-subject-out (LOSO)
scheme was used on the FG-NET database. Each
strong age classifier was built with only 200 weak
classifiers, as the LOSO scheme costs a lot of time.
This comparison was conducted using many other age
estimation methods, including (Guo et al., 2009b)’s
BIF or (Yan et al., 2007)’s RUN. The best reported
result isobtained by (Luu et al., 2011) with a mean
absolute error (MAE) value of 4.12 years. Other re-
ports detailed results on several age subsets as shown
in table 1. We obtain results close to Guo’s perfor-
mance (Guo et al., 2009b) with an MAE value of 4.94
years (0.04 year difference with human perception er-
ror). Our results are the best available on the [0-9] and
[40+] subsets and the second best on the [20-39] sub-
set. Cumulative scores on FG-NET are available in
figure 2. Table 1 compares mean errors for each age
subset. The best age estimation methods (including
ours) have results close to human perception on this
database.
Human perception errors and estimation errors on
the 0 to 100 database are presented in figure 3. This
is pure generalization as our estimators were trained
on FG-NET and then tested on this database. Our es-
timator was not trained for age boundaries above 60
years. It explains its poor performance on this gener-
0 2 4 6 8 10 12 14 16 18 20
0
10
20
30
40
50
60
70
80
90
100
Cumulative Score (%)
Error Level (years)
Accuracy vs. Error Level (FG−NET)
Boosted LBP hist.
BIF
QM
BM
MLP
Human estimation on 60 images subset
Figure 2: Human perception and estimation errors: compar-
ison with state-of-art works (FGNET database).
0 5 10 15 20 25 30 35 40 45 50
0
20
40
60
80
100
Error level (years)
CumulativeScore (%)
Generalization test on 0 to 100 video
Humans 0 to 60
Humans 0 to 100
Automatic Estimator 0 to 60
Automatic Estimator 0 to 100
Figure 3: Human perception and estimation errors (” 0 to
100” database).
alization experiment over the full ”0 to 100” database
and its acceptable results (not far from human opera-
tors) on the ”0 to 60” subset.
5.3 Results on Sequences
In the following experiments, the complete system is
tested on our video dataset. The cropped images and
LBP histogram signatures used by both gender and
age estimators are the same. Each frame is computed
to make a complete observation of the estimation out-
puts.
Gender Estimation. For the video set, our gender
estimator outputs distributions of real values between
0 and 1. Using E-M algorithm, each video’s most
weighted distributions’ mean is on the correct side of
the 0.5 threshold. As the estimation score is centered
on 0.5, we can study the estimator’s outputs distribu-
tion considering several rejection thresholds δ. For in-
stance, with δ = 0.01, outputs within the [0.49 0.51]
interval are discarded. The ROC-like graphs shown
in figure 4 are plotted using the subsequent good and
false response rates. This experiment uses two set-
tings of regions of interest for the LBP histograms
(figure 4.a). Both settings use regions of interest of
the same size,
1
5
of the original 128 × 96 crop. The
first setting uses 25 (5 × 5) non-overlapping 26 × 20
pixels regions across the face crop and the second
uses 64 (8 × 8) 26 × 20 pixels regions. The 8 × 8
blocks setting performs slightly better on the lateral
illumination dataset (figure 4.b). The use of symme-
LiveStreamOrientedAgeandGenderEstimationusingBoostedLBPHistogramsComparisons
795
Table 1: Age estimation: results on the FG-NET database, using LOSO scheme test. Mean absolute error in years, on every
age subsets. QM and MLP methods are described in (Lanitis et al., 2004).
Range #img. Ours BIF RUN QM MLP
0-9 371 2.42 2.99 2.51 6.26 11.63
10-19 339 3.92 3.39 3.76 5.85 3.33
20-29 144 4.95 4.30 6.38 7.10 8.81
30-39 70 10.56 8.24 12.51 11.56 18.46
40-49 46 14.59 14.98 20.09 14.80 27.98
50-59 15 18.45 20.49 28.07 24.27 49.13
60-69 8 27.62 31.62 42.50 37.38 49.13
0-69 1002 4.94 4.77 5.78 5.57 10.39
0 10 20 30 40 50 60 70 80 90 100
86
88
90
92
94
96
98
100
gender estimator behaviour on all video sequences
classification score (%)
rejection rate (%)
5x5 Blocks, No Sym., AUC=0.9844
5x5 Blocks, Sym., AUC=0.9853
8x8 Blocks, No Sym., AUC=0.9844
8x8 Blocks, Sym., AUC=0.9858
Pixel Comp., No Sym., AUC=0.9696
(a)
0 10 20 30 40 50 60 70 80 90 100
86
88
90
92
94
96
98
100
gender estimator behaviour on video sequences with lateral illumination
classification score (%)
rejection rate (%)
5x5 Blocks, No Sym., AUC=0.9772
5x5 Blocks, Sym., AUC=0.9795
8x8 Blocks, No Sym., AUC=0.9822
8x8 Blocks, Sym., AUC=0.9828
Pixel Comp., No Sym., AUC=0.969
(b)
Figure 4: Gender estimator behavior and comparison with
Baluja’s estimator (Baluja and Rowley, 2007) for different
settings (a) and illuminations (b).
try refers to using each source frame’s symmetric im-
age, thus using two MBLBP histogram signatures for
each computed frame. As the learning database itself
was mirrored, using symmetry does not dramatically
improves the results. The pixel comparisons method
is an implementation of Baluja’s boosted gender es-
timator, using 30 × 30 face images. It was used as a
baseline for our latest experiments.
Age Estimation. Accordingly to the previous results,
we choose to use the (8 × 8) blocks settings. Age es-
timation results over the video set are shown as cu-
mulative scores in figure 5: 65.4% of the estimations
are within the 5 years threshold on the neutral illu-
mination set. Despite having mirrored the database,
we still obtain slightly better results on this dataset.
These results are comparable to the human age per-
ception experiment results on FG-NET or the 0 to 60
dataset seen in figure 3.
0 2 4 6 8 10 12 14 16 18 20
0
20
40
60
80
100
X: 5
Y: 65.41
X: 10
Y: 81.7
Error level (years)
Cumulative score (%)
Cumulative score on video set
Neutral illumination
Lateral illumination
All sequences
Figure 5: Age estimation: cumulative scores on the video
set with neutral or with added lateral illumination.
6 CONCLUSIONS AND
PERSPECTIVES
An alternative method for live stream oriented age and
gender estimation is provided. It uses boosted com-
parisons over uniform LBP histograms based facial
signatures. The system provides real-time estimations
and is able to track several targets simultaneously.
The system’s performance is compared favorably to
state of the art techniques of age and gender recog-
nition on common databases. Other experimentations
are conducted on video sequences and on live streams
to show the accuracy of the whole process (figures 6
and 7: tracking, pose, gender and age estimation).
Apart from the inevitable database collection
needed to improve our training sessions, many per-
spectives appear. The next natural evolution would
be to change the face cropping for face warping using
our 3D model to be more resilient to face orientation,
instead of only rejecting extreme poses. The present
final age and gender decision uses simple logic over
the strong classifiers. It would be interesting to build
a system able to decide from all the strong classifiers
outputs.
ACKNOWLEDGEMENTS
We provide our test videos including the ground truth
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
796
Figure 6: Results in outdoor conditions: face tracking, pose
estimation, age and gender estimation (test video n
o
1).
measures on request. Please contact us by mail to
receive our data. The authors gratefully acknowl-
edge the contribution of the Agence National de la
Recherche (CIFRE N533/2009).
Figure 7: Results in outdoor conditions: face tracking, pose
estimation, age and gender estimation (test video n
o
2).
REFERENCES
Anastasi, J. and Rhodes, M. (2005). An own-age bias in
face recognition for children and older adults. Psy.
Bul. & Rev., 12(6):1043–1047.
Baluja, S. and Rowley, H. (2007). Boosting sex identifica-
tion performance. IJCV, 71:111–119.
LiveStreamOrientedAgeandGenderEstimationusingBoostedLBPHistogramsComparisons
797
Cottrell, G. and Metclafe, J. (1990). Empath: face, emotion,
and gender recognition using holons. Proc. of Adv. in
NIPS, 3:567–571.
Freund, Y. and Schapire, H. (1996). Experiments with a new
boosting algorithm. In Int. Conf. on Machine Learn-
ing, pages 148–156.
Fu, Y., Guo, G., and Huang, T. (2010). Age synthesis and
estimation via faces: A survey. IEEE Trans. on PAMI,
32(11):1955–1976.
Gao, F. and Ai, H. (2009). Face age classification on
consumer images with gabor feature and fuzzy lda
method. LNCS, 5558:132–141.
Golomb, B., Lawrence, D., and Sejnowski, T. (1991).
Sexnet, a neural network identifies sex from human
faces. NIPS, 3.
Guo, G., Mu, G., Dyer, D., and T.S., H. (2009a). A study
on automatic age estimation using a large database.
ICCV, pages 1986–1991.
Guo, G., Mu, G., Fu, Y., and Huang, T. (2009b). Human age
estimation using bio inspired features. CVPR, pages
112–119.
Hadid, A. and M., P. (2008). Combining motion and appear-
ance for gender classification from video sequences.
ICPR, pages 1–4.
Han, H., Otto, C., and Jain, A. (2013). Age estimation from
face images: Human vs. machine performance. In
Proc. ICB, pages 4–7.
Huang, G., Ramesh, M., Berg, T., and Learned-Miller, E.
(2007). Labeled faces in the wild: A database for
studying face recognition in unconstrained environ-
ments. Uni. of Massachusetts, Tech. Report 07-49.
Kumar, N., Belhumeur, P., and Nayar, S. (2008). Facetracer:
A search engine for large collections of images with
faces. ECCV, pages 340–353.
Lanitis, A., Draganova, C., and Christodoulou, C. (2004).
Comparing different classiers for automatic age esti-
mation. IEEE Trans. on SMC-B, 34(1):621–628.
Li, X., Maybank, S., Yan, S., Tao, D., and D., X.
(2008). Gait components and their application to gen-
der recognition. IEEE Trans. on SMC-C, 38(2):145–
155.
Liao, S., Zhu, X., Lei, Z., Zhang, L., and Li, S. (2007).
Learning multi-scale block local binary patterns for
face recognition. ICB, pages 828–837.
Luu, K., Ricanek, K., Bui, T., and Suen, C. (2009). Age es-
timation using active appearance models and support
vector machine regression. the IEEE Conf. on Bio-
metrics: Theory, Applications, and Systems (BTAS),
pages 1–5.
Luu, K., Seshadri, K., Savvides, M., and Bui, T.D.and Suen,
C. (2011). Contourlet appearance model for facial
age estimation. Int. Joint Conf. on Biometrics (IJCB),
pages 1–8.
Makinen, E. and Raisamo, R. (2008). An experimental
comparison of gender classification methods. Pattern
Recognition Letters, 29(10):1544–1556.
Moghaddam, B. and Yang, M. (2002). Learning gender
with support faces. IEEE trans. on PAMI, 24(5):707–
711.
Ojala, T., Pietikinen, M., and Maenpaa, T. (2002). Multires-
olution gray-scale and rotation invariant texture clas-
sification with local binary patterns. IEEE Trans. on
PAMI, 24:7:971–9.
Phothisane, P., Bigorgne, E., Collot, L., and Prevost, L.
(2011). A robust composite metric for head pose
tracking using an accurate face model. In Proc. FG,
pages 694–699.
Ramanathan, N. and Chellappa, R. (2006). Face verifica-
tion across age progression. IEEE Trans. on Image
Processing, 15(11):3349–3361.
Shakhnarovich, G., Viola, P., and Moghaddam, B. (2002).
A unified learning framework for real time face detec-
tion and classification. FG, pages 14–21.
Shan, C. (2012). Learning local binary patterns for gender
classification on real-world face images. Patt. Recog.
Letters, 33(4):431–437.
Shan, C., Gong, S., and McOwan, P. (2008). Fusing gait
and face cues for human gender recognition. NVR,
71(10-12):1931–1938.
Thukral, P., Mitra, K., and Chellappa, R. (2012). A hierar-
chical approach for human age estimation. ICASSP,
pages 1529–1532.
Wild, H., Barett, S., Spence, M. J., O’Toole, A., Cheng,
Y., and Brooke, J. (2000). Recognition and sex cate-
gorization of adults’ and children’s faces: Examining
performance in the absence of sex-stereotyped cues.
Jour. of Exp. Child Psychology, 77:269–291.
Yan, S., Wang, H., Tang, X., and T.S., H. (2007). Learning
auto-structured regressor from uncertain nonnegative
labels. ICCV, pages 1–8.
Zhang, W., Shan, S., Zhang, H., Gao, W., and Chen, X.
(2005). Multi-resolution histograms of local vari-
ation patterns (mhlvp) for robust face recognition.
Audio- and Video-Based Biometric Person Authenti-
cation, pages 7:937–944.
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
798
