automatic facial age estimationyunfu/papers/pricai10_t4.pdf · model age progression in young faces...
TRANSCRIPT
Xin Geng1,2
, Yun
Fu3
, Kate Smith‐Miles1
Automatic Facial Age Estimation
1
Monash
University, Australia2
Southeast University, China3
University at Buffalo (SUNY), USA
Tutorial at PRICAI 2010
2:00pm-5:00pm, Monday, August 30, 2010
•
Introduction•
Overview of existing techniques
•
Evaluation of existing techniques•
Our approaches
•
Conclusions and discussions
Outline
•
Introduction–
Human facial aging
–
What is facial age estimation?–
Why we need the techniques for automatic facial age
estimation?–
Main challenges
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Outline
•
Introduction–
Human facial aging
–
What is facial age estimation?–
Why we need the techniques for automatic facial age
estimation?–
Main challenges
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Outline
Who is this man?
Human Facial Aging
Facial aging of Albert EinsteinFacial aging of Albert Einstein
With the progress of age, the appearance of human faces exhibits remarkable changes
•
Craniofacial growth–
Face contour
–
Facial feature (eyes, nose, mouth etc.) shape–
Facial feature distribution on face
•
Skin aging–
Skin color
–
Wrinkles•
Reduction of muscle strength/ elasticity–
Facial lines
–
Wrinkles–
Facial feature shape
–
Facial feature distribution on face
Why?
And why?•
2009 Nobel Prize in Physiology or Medicine
Elizabeth H. Blackburn Carol W. Greider Jack W. Szostak
Discover the relationship of telomeres and the aging of cells
•
Slow•
Irreversible
•
Personalized–
Internal influential factors•
Gene
•
Gender–
External influential factors•
Health
•
Living style•
Working environment
•
Sociality•
Weather conditions
•
Smoking•
…
Characteristics of the Aging Progress
•
Early age (from birth to adulthood)–
Mainly shape changes caused by craniofacial growth•
Forehead Slopes back, shrinks and releases spaces on the surface of the
cranium •
Facial featuresExpand their areas and tend to cover the interstitial spaces
•
CheeksExtend to larger areas
•
ChinBecomes more protrusive
Facial Aging Stages
•
Early age (from birth to adulthood)–
Mainly shape changes caused by craniofacial growth
Facial Aging Stages
[Ramanathan and Chellappa, CVPR’06]
[Todd et al., Scientific American, 1980]
•
Early age (from birth to adulthood)–
Minor skin changes•
Facial hairs Become dense, change color
•
Skin colorSlightly changes
Facial Aging Stages
Facial Aging Stages•
During the adulthood (from adulthood to old age)–
Mainly skin changes (texture changes)•
Thinner
•
Darker•
Less elastic
•
More leathery
•
Adynamic/ Dynamic wrinkles•
Blemishes
•
Double chin•
Dropping cheek
•
Eyelid bags …
[Gonzalez-Ulloa and Flores, Plast. Reconstr. Surg. 1965]
Facial Aging Stages•
During the adulthood (from adulthood to old age)–
Minor craniofacial growth•
Face shape–
U‐shaped
–
Upside‐down triangle–
Trapezoid
–
Rectangle
Outline•
Introduction–
Human facial aging
–
What is facial age estimation?–
Why we need the techniques for automatic facial age
estimation?–
Main challenges
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Facial Age Estimation•
Definition
Label a face image automatically with the exact age (year) or the age group (year range) of the individual face
Facial Age Estimation Age = 3 (years)
Facial Age Estimation•
Reversely…–
Facial Age Synthesis
Re‐render a face image aesthetically with natural aging or rejuvenating effects on the individual face.
Facial Age SynthesisInput Image
Synthesized Image at the Target Age
Target Age (Older)
Facial Age Estimation•
About the Age–
Actual Age (Chronological Age)
The number of years an individual has lived, i.e., “the real age”.
–
Appearance Age The age information revealed by the visual appearance.
–
Perceived Age The individual age gauged by human subjects from the
visual appearance.–
Estimated Age
The indiviaul age recognized by machine from the visual appearance.
Ground truth
Usually consistent with the actual age, but variation is often inevitable
Defined on visual appearance. The closer to the appearance age, the better.
Which is better?
Outline•
Introduction–
Human facial aging
–
What is facial age estimation?–
Why we need the techniques for automatic facial age
estimation?–
Main challenges
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Motivation•
Age estimation is a basic ability required for smooth
communication between humans. –
People at different ages have different requirements and
preferences in various aspects, such as •
Linguistics
•
Aesthetics•
Consumption habit
•
…–
A good understanding of age leads to successful
communication
between humans•
What about human‐computer interaction?
Motivation•
Human ability in age estimation–
Developed early in life
–
Can be fairly accurate–
But can be affected by various factors•
More accurate in the estimator’s own kind–
Age group
–
Race–
Gender
•
Facial attachment–
Glasses
–
Beard•
Attractiveness
•
Kindness
Motivation•
Is it easy?
–
Difficult for humans!–
Even more difficult for machines!
19 49 5
Motivation•
But it is very useful!–
Age‐specific HCISatisfy preferences of all ages
–
Age‐specific access control Children protection
–
Law enforcement Security and Surveillance
–
Multi‐cue identification face/fingerprint/iris + age
Applications•
e.g., Internet safety for minors
Applications•
e.g., Cigarette vending machine
Applications•
e.g., Age‐specific shopping HCI
Outline•
Introduction–
Human facial aging
–
What is facial age estimation?–
Why we need the techniques for automatic facial age
estimation?–
Main challenges
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Challenges –
Incomplete aging patterns•
The aging progress is uncontrollable–
No one can age at will
–
The procedure of aging is slow and irreversible–
The collection of sufficient training data for age
estimation is extremely laborious•
Consequently, –
The available dataset typically just contain a very
limited number of aging images for each person–
The images at the higher ages are especially rare
Challenges ‐
personalized aging patterns•
Different persons age in different ways–
Internal influential factors•
Gene
•
Gender–
External influential factors•
Health
•
Living style•
Working environment
•
Sociality•
Weather conditions
•
Smoking•
…
•
Consequently, –
The mapping from features (face images) to class labels (ages) is
not unique
Challenges ‐
temporal aging patterns•
The aging progress must obey the order of time. –
The face status at a particular age will affect all older faces,
but will not affect those younger ones.•
Consequently, –
The set of class labels (ages) is a totally ordered set
–
Each age has a unique rank in the time sequence
Outline•
Introduction
•
Overview of existing techniques–
Aging face models
–
Age estimation algorithms•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Flow chart of age estimation systems
Aging face model
Age estimation algorithms
Face image Age
Outline•
Introduction
•
Overview of existing techniques–
Aging face models
–
Age estimation algorithms•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Aging face models•
Anthropometric models–
Based on the measurements and proportions of the human faces
•
Active appearance models–
Based on the statistical face model AAM proposed by T. Cootes
et al.
•
Aging pattern subspace–
Based on the AGES method proposed by Geng
et al.
•
Age manifold–
Based on the manifold embedding techniques to learn the low‐
dimensional aging trend•
Appearance feature models–
Based on aging‐related features extracted from face images
Anthropometric models•
The earliest work on age estimation based on face
images –
[Kwon and Lobo, CVPR’94] [Kwon and Lobo, CVIU’99]
–
Six
ratios of distances on frontal face images to separate babies from adults
Anthropometric models•
Face recognition across age progression –
[Ramanathan
and Chellappa, CVPR 2006] –
Model age progression in young faces with eight
ratios of
distance measures to predict an individual’s appearance across different ages
Eight ratiosEight ratios
(1) Facial index
(2) Mandibular index
(3) Intercanthal index
(4) Orbital width index
(5) Eye fissure index
(6) Nasal index
(7) Vermilion height index
(8) Mouth-Face width index
Anthropometric models•
Might be useful to distinguish minors from adults,
but not appropriate for classifying different adult ages
–
E.g., Kwon and Lobo [ CVPR’94, CVIU’99]
further analyze the wrinkles to separate young adults from senior adults
Anthropometric models•
Sensitive to head pose–
Usually only frontal faces are used to measure the facial
geometries–
All tested on small data sets
•
Do not involve texture information
Active appearance models•
The appearance model [Edwards et al., IVC’98]–
Combine a model of shape
variation with a model of
texture
variation–
Require a training set of annotated images where the
‘landmark points’
have been marked on each example
Active appearance models•
The appearance model [Edwards et al., IVC’98]–
Shape•
The face shape in each image is represented as a vector x
•
Procrustes
analysis is used to align the shape vectors over the training set
•
Apply PCA to the shape vectors x
to get the shape model–
Texture•
Warp each training image so the points match those of the mean
shape, obtaining a “shape‐free patch”•
A texture vector g
is created via raster scan, and normalized
•
Apply PCA to the texture vectors g
to get the texture model
Active appearance models•
The appearance model [Edwards et al., IVC’98]–
Combine shape and texture models
•
: the mean shape•
: the mean texture
•
: matrices describing the modes of variation derived from the training set
•
: the appearance model parameters•
How? –
concatenate the shape model parameters and
texture model parameters and apply PCA again
Active appearance models•
The active appearance model [Cootes, et al., TPAMI’01]–
The appearance model relies on the ‘landmark points’
–
In the training set, the landmark points are manually labeled
–
Given a new image, an active search process
is required to find the parameters that make the appearance model matches the image as closely as possible
Active appearance models•
Age estimation based on AAM–
Aging function:
age
= f(b)
[Lanitis
et al. TPAMI’
02]
•
b: the vector of the AAM parameters•
f: a quadratic function.
•
Training–
f() is fitted by Genetic Algorithm (GA) for each individual in the
training set as his/her aging function•
Age estimation
–
Substitute the AAM parameters of a new face image into the aging
function. The output of the function is estimated age–
Other variations based on aging function
[Lanitis
et al. TSMCB’04]
–
AGES also uses AAM [Geng
et al. MM’06], [Geng
et al. TPAMI’07]
Active appearance models
•
AAM vs.
Anthropometric model–
AAM considers both the shape
and texture, while
the anthropometric model only involve facial geometry
–
AAM based approaches can deal with any age, while the anthropometric model can be only used
to distinguish minors from adults–
AAM is robust against head poses, while the
anthropometric model is quite sensitive to poses
Aging pattern subspace•
Regard the sequence of an individual’s aging face
images as a whole
rather than separately•
Based on a data structure called aging pattern
•
Two stages: learning stage & age estimation stage•
Learning stage –
Model the aging patterns by constructing a representative
subspace•
Age estimation stage–
Find the most suitable age for the single test face image
based on the subspace of image sequences•
Refs: the AGES algorithm series
[Geng
et al., MM’06], [Geng
et al., TPAMI’07], [Geng
et al., MM’08], [Geng
et al., ICASSP’09]
Aging Pattern
An aging pattern is a sequence of personal face images sorted in time order
Appearance Model[Edwards et al. IVC’98]
Incomplete aging pattern causes many missing values in the aging pattern vector
Aging Pattern•
Advantages
of using aging pattern as basic data sample
–
Each aging pattern involves only one person –
personalized–
All the images in one aging pattern are sorted in the time order
–
temporal
•
New challenges–
The aging patterns are always incomplete•
There are many missing values
in the aging pattern vector
–
An aging pattern is an image sequence, which corresponds to a class (age) sequence
•
But what we need finally is to give a single age estimation
to a single face image
Aging pattern subspace•
AGESAGing
pattErn
Subspace
•
The goalLearn a representative subspace for the aging patterns
•
How to deal with the massive missing values?–
Apply PCA iteratively
•
How to predict the age of a single face image based on the subspace of image sequences?
–
Put the face image into different positions in the aging pattern
–
Find the position with minimum reconstruction error
Aging pattern subspace•
Well match the characteristics of the age estimation
problem•
High request for the quality of training samples, but
with certain ability to handle missing face images in the aging sequence
•
Extensibility–
Feature extraction methods (AAM or any other?)
–
Linear PCA Nonlinear subspace analysis Multilinear
(tensor) subspace analysis
…
Age manifold•
Assume that the aging face images are distributed on an
intrinsic low‐dimensional manifold (faces with close ages locate closely on the manifold)
•
Regard the age estimation problem as a special case of supervised manifold embedding
problem
•
Rely on the power of manifold learning to find out the mapping from the face images to the ages
Age manifold•
Problem formulation
(for any supervised manifold
learning, not specified for age)–
Given•
Samples:
•
Class labels:–
The goal•
Find the low‐dimensional embeddings
•
The mapping from the original space to the manifold space
Age manifold•
Typical approaches–
Orthogonal Locality Preserving Projections (OLPP)
[Cai
et al., TIP’06]–
Conformal Embedding Analysis (CEA)
[Fu and Huang, TMM’08]–
Locally Adjusted Robust Regression (LARR)
[Guo
et al., TIP’08], [Guo
et al., WACV’08]–
Synchronized Submanifold
Embedding (SSE)
[Yan et al., TIP’09]
Age manifold•
Could find the intrinsic low‐dimensional manifold for
the aging face images•
Do not require many images at different ages from
the same person•
But require many images labeled with age to learn
the embedded manifold with statistical sufficiency•
Usually not tailored to the characteristics of the age
estimation problem •
Can be applied to other problems as well
Appearance feature models•
Extract age‐related features based on the face
appearance–
Texture features vs. Shape features
–
Global features vs. Local features–
Graphics / image / signal processing features
–
Biologically inspired features
Appearance feature models•
Typical approaches–
[Hayashi et al., ICVSM’01, SICE’02]•
Consider both texture (wrinkle) and shape (geometry)
features•
A semantic‐level indexing of the face was also used
•
Extended by [Fujiwara et al., KES’03]–
[Fukai
et al., SICE’07]
•
Use Fast Fourier Transform (FFT) to extract feature spectrum
•
Use Genetic Algorithm (GA) for feature selection
Appearance feature models•
Typical approaches–
[Günay
and Nabiyev, ISCI’08]
•
Rely on the texture descriptor Local Binary Patterns (LBP)
–
[Gao
and Ai, ICAB’09]•
Rely on the Gabor features
–
[Yan et al., CVPR’08, ICASSP’08]•
Use Spatially Flexible Patch (SFP) as local feature
descriptor
Appearance feature models•
Typical approaches–
[Suo
et al., FGR’08]
•
Use four types of features: topology, geometry, photometry and configuration
•
Hierachical
face model ‐
a face image is decomposed into detailed parts from coarse to fine
–
[Guo
et al., CVPR’09]•
Use Biologically Inspired Features (BIF)
–
Based on the ‘HMAX’
model – a feed‐forward model of the primate
visual object recognition pathway–
[Guo
et al., ICCV’09]
•
BIF + age manifold features
Outline•
Introduction
•
Overview of existing techniques–
Aging face models
–
Age estimation algorithms•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Age estimation algorithms•
Closely related to the aging face models used–
Strong facial feature + simple age estimation algorithm
–
Simple facial feature + strong age estimation algorithm–
Integrated aging face model and age estimation
algorithm•
Age estimation is a special PR problem–
Each age is a class label –
classification
–
Each age is a number –
regression–
Combine classification and regression?
–
Beyond classical classification and regression?
Classification•
[Lanitis
et al., TSMC‐B’04]
–
General purpose classification methods applied to age estimation
•
Nearest neighbor•
Multilayer perceptron
•
Self‐organizing map (SOM)•
[Ueki et al., FGR’06]–
Age group classification
–
Two‐phase feature extraction: 2DLDA+LDA–
One Gaussian model is trained for each age group
–
Classification is based on likelihoods of the Gaussian models
Classification•
[Guo
et al., WACV’08, TIP’08]
–
Apply SVM to age estimation•
[Kanno
et al., IEICE TIS’01]
–
Artificial neural network–
4‐class age‐group classification
•
AGES [Geng
et al., MM’06], [Geng
et al., TPAMI’07], [Geng
et al., MM’08], [Geng
et al., ICASSP’09]
–
Age classification based on aging pattern subspace–
Compare AGES with general purpose classification
methods: kNN, BP, C4.5, SVM
Regression•
Aging function
[Lanitis
et al. TPAMI’
02, TSMCB’04]–
Three formulations for AF: linear, quadratic, and cubic
–
Output of the AF is the estimated age•
Multiple linear regression
[Fu et al., ICME’07, TMM’08]–
Fit the CEA age manifold
•
Support Vector Regression (SVR) [Guo
et al., WACV’08, TIP’08]
–
Fit the OLPP age manifold
Regression•
[Yan et al ., ICCV’07]–
Perform age estimation via a regressor
learned from
uncertain nonnegative labels–
Use Semi‐Definite Programming (SDP) to solve the
regression problem•
Image Based Regression (IBR)
[Zhou et al., ICCV’05]–
A boosting scheme is used to select features from
redundant Haar‐like feature set•
Robust Multi‐instance Regression (RMIR)
[Ni et al., MM’09]–
Automatic online training
Hybrid Approach•
Locally Adjusted Robust Regression (LARR)
[Guo
et al., WACV’08, TIP’08]–
Global regression by SVR
–
Locally adjust the estimated age via classification (SVM)–
Local search range is determined heuristically
•
Further improvement of LARR [Guo
et al., SVPR‐SLAM’08]
–
Determine the combination parameters automatically •
Transform both the regression (SVR) and classification
(SVM) results into probabilities using a uniform distribution
•
Combine the results via Bayesian theory
Beyond classical classification and regression?•
Classification–
What if the class labels are closely related to each other?
–
What if the training samples for different classes are unbalanced?
•
Regression–
How to constrain the output within a desired range?
–
How to choose a suitable model for regression?•
An approach beyond all these –
Learning from Label Distributions (LLD) [Geng
et al., AAAI’10]
–
Particularly useful when•
The classes are correlated to each other
•
The training data for some classes are insufficient
Matching the problem
•
Neighboring ages are highly correlated–
Aging is a slow and gradual progress
–
The faces at close ages look quite similar•
The ‘lack of training samples’
problem
–
The available dataset typically just contain a very limited number of aging images for each person
–
The images at the higher ages are especially rare•
LLD matches the problem of age estimation well–
Use the neighboring ages to relieve the ‘lack of
training samples’
problem
Label distribution
•
A real number is assigned to each class label•
Those numbers for all the labels sum to 1
•
Similar to but NOT probability
LLD for age estimation•
Each image is assigned with a label distribution–
Highest at the real age
–
Decrease with the distance from the real age–
E.g., Gaussian distribution, triangle distribution
•
When learning a particular age, the neighboring ages are also utilized
–
The contribution is reversely proportional to the distance from the neighboring age to the real age
LLD for age estimation•
Learning from label distributions–
Learn a conditional p.d.f. as similar as possible to the label
distribution –
Similarity is measured by the Kullback‐Leibler
divergence
•
Age estimation–
Given a face image x
–
Calculate its label distribution–
Explicit label distribution of x provides many possibilities in classifier
design•
Single label
•
Multiple labels•
Prediction with confidence measure
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques–
Aging face databases
–
Evaluation metrics–
Comparison of existing techniques
•
Our approaches•
Conclusions and discussions
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques–
Aging face databases
–
Evaluation metrics–
Comparison of existing techniques
•
Our approaches•
Conclusions and discussions
Aging face databases•
FG‐NET Aging Database–
1002 face images
–
82 subjects–
Age: 0‐69
–
Publicly available•
YGA Database–
8000 face images
–
1600 subjects–
Age: 0‐93
Aging face databases•
MORPH Database–
Album 1•
1724 face images
•
515 subjects•
Age: 15‐68
–
Album 2•
> 20,000 face images
•
>4000 subjects•
Age: 21‐99
–
Publicly available
Aging face databases•
WIT‐DB Database–
26,222 face images
–
5500 subjects–
Age: 3‐85
•
AI&R (V2.0) Asian Face Database–
34 face images
–
17 subjects–
Age: 22‐61
•
Burt’s Caucasian Face Database–
147 face images
–
147 subjects–
Age: 20‐62
Aging face databases•
LHI Face Database–
8,000 face images
–
8,000 subjects–
Age: 9‐89
•
HOIP Face Database–
306,000 face images
–
300 subjects–
Age: 15‐64
•
Iranian Face Database–
3,600 face images
–
616 subjects–
Age: 2‐85
Aging face databases•
Gallagher’s Web Collected Database–
28,231 faces
–
5,080 images–
Seven age ranges: 0‐2, 3‐7, 8‐12, 13‐19, 20‐36, 37‐65, 66+
•
Ni’s Web Collected Database–
219,892 faces
–
77,021 images–
Age: 1‐80
•
3D Morphable
Database–
200 adults
–
238 teenager
Aging face databases•
Summary–
Publicly available•
FG‐NET Aging Database
•
MORPH Database•
Gallagher’s Web Collected Database
–
Large size•
MORPH Database
•
YGA Database•
LHI Face Database
•
NI’s
Web Collected Database•
Gallagher’s Web Collected Database
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques–
Aging face databases
–
Evaluation protocols–
Comparison of existing techniques
•
Our approaches•
Conclusions and discussions
Evaluation protocols•
Two measurements–
Mean Absolute Error (MAE)
•
First used in [Lanitis
et al., TPAMI’02]•
An indicator of the average performance of the
age estimators–
Cumulative Score (CS) at error level l
•
First used in [Geng
et al., MM’06]•
An indicator of accuracy of the age estimators
Evaluation protocols•
An example of CS [Geng
et al., MM’06]
Evaluation protocols•
Training Set vs. Test Set–
Half‐half
–
Cross validation–
Leave One Person Out (LOPO)
•
Suggested evaluation protocol–
MAE + Cumulative Score
–
LOPO
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques–
Aging face databases
–
Evaluation protocols–
Comparison of existing techniques
•
Our approaches•
Conclusions and discussions
Comparison of existing techniques•
Summary table [Fu et al., TPAMI’10]
C
Comparison of existing techniques•
Summary table [Fu et al., TPAMI’10]
C
Comparison of existing techniques•
Summary table [Fu et al., TPAMI’10]
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches–
AGES
–
Learning from label distributions•
Conclusions and discussions
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches–
AGES
–
Learning from label distributions•
Conclusions and discussions
AGES Series•
The original
1.
Xin
Geng, Zhi‐Hua
Zhou, Yu Zhang, Gang Li, and Honghua
Dai. Learning from
facial aging patterns for automatic age estimation. ACM MM'06, Santa Barbara,
CA, 2006, pp. 307‐316. 2.
Xin
Geng, Zhi‐Hua
Zhou, and Kate Smith‐Miles. Automatic age estimation
based on facial aging patterns. IEEE TPAMI, 2007, 29(12): 2234‐2240.
•
Variants–
Nonlinear3.
Xin
Geng, Kate Smith‐Miles, and Z.‐H. Zhou. Facial Age Estimation by
Nonlinear Aging Pattern Subspace. ACM MM'08, Vancouver, Canada, 2008, pp.
721‐724.
–
Multilinear4.
Xin
Geng
and Kate Smith‐Miles. Facial Age Estimation by Multilinear
Subspace
Analysis. ICASSP’09, Taipei, Taiwan, 2009, pp. 865‐868.
Aging patternFace image
Iterative learning algorithm
AGES
Incomplete temporal aging patterns which are determined by personalized factors
Difficulties
& Countermeasures
Aging Pattern
An aging pattern is a sequence of personal face images sorted in time order
Appearance Model[Edwards et al. IVC’98]
Aging Pattern•
Advantages &
New Challenges
Aging Pattern
Identity
Time
Personalized
Temporal
Difficulties
Incomplete
Image sequence
class sequence
AGES – Learning•AGES
AGing
pattErn
Subspace•The goal
Learn a representative subspace for the aging patternsPCA (Principal Component Analysis)?
•Basic idea
Different person Same age
Same person Different ages+
Missing faces
?
mm
mm
m m
m mm
age
person
AGES – LearningApply PCA iteratively on the incomplete aging patterns
Aging Patterns
Subspace
PCA Reconstruction
Improved
Improved
AGES – Learning
Initialization (including the first PCA)
Reconstruction
PCA
AGES – Learning
0 2 4 6 8
10
12 14 16 18
Full-fill the aging patterns
AGES – Age Estimation
LDA+AGES•
There are usually variations other than aging in the
data set•
LDA can be applied to the feature vectors with age
labels•
The resulting discriminant
parameters are expected
to be more related to the aging variation•
The AGES based on such discriminant
parameters is
denoted by AGESlda
Experiments•
Data set: FG‐NET Aging Database–
1002
face images
–
82
subjects–
6‐18
images each
subject–
Age: 0‐69
–
Variations: illumination, pose,
expression, beards, moustaches, spectacles, hats
Experiments•
Data set: MORPH Database–
433
face images
–
Around
3
images each subject–
Age: 16‐68
–
Used to test the algorithms trained on the FG‐NET database
Experiments•
Compared Methods–
AGES
–
AGESlda
–
WAS, [Lanitis
et al. TPAMI, 2002]–
AAS, [Lanitis
et al. TSMCB, 2004]
–
Conventional classification methods•
kNN
•
BP network•
C4.5 decision tree
•
SVM
Experiments•
Compared Methods–
Human observers (29 college students)•
HumanA
test
•
HumanB
test
Experiments•
Compare Mode–
For AGES, AGESlda
,
WAS, AAS, kNN, BP, C4.5, SVM•
FG‐NET:
Leave‐One‐Person‐Out (LOPO)
•
Train on FG‐NET and test on MORPH–
For human tests5% face images randomly selected from the FG‐NET
Aging Database ( )
•
Evaluation measure–
MAE + Cumulative score
1002 5% 51× ≈
Result – Age Estimation•
MAE
Result – Age Estimation•
Cumulative Scores
FG-NET (LOPO) MORPH (Test Set)
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches–
AGES
–
Learning from label distributions•
Conclusions and discussions
Learning from label distributions•
Xin
Geng, Kate Smith‐Miles, Zhi‐Hua
Zhou. Facial
Age Estimation by Learning from Label Distributions. AAAI’10, Atlanta, GA, 2010.
The ‘lack of training samples’
problem•
Ideal training set should–
Cover a wide range of age
–
Include many different subjects–
Contain at least one image for each age of each subject
•
Unfortunately…–
The aging progress cannot be artificially controlled
–
Great efforts in searching for the images taken years ago–
Can do nothing to acquire future images
•
Consequently, –
The available dataset typically just contain a very limited number
of aging images for each person–
The images at the higher ages are especially rare
The neighboring ages
•
Aging is a slow and gradual progress•
The faces at close ages look quite similar
•
Can we use the neighboring ages to relieve the ‘lack of training samples’
problem?
Label distribution
•
A real number is assigned to each class label y, representing the degree that the label
describes the instance•
Those numbers for all the labels sum to 1
•
Similar to but NOT probability
Special cases of label distribution
•
Case 1: Single label•
Case 2: Multiple labels
•
Case 3: General case of label distribution
Learning from label distributions•
Problem definition–
Training set
–
Goal: learn a conditional p.d.f. from S
•
The best parameter vector
The Kullback‐Leibler divergence from to
Learning from label distributions•
The maximum entropy model (Berger, et al. 1996)
•
Target function of for the optimization:
,exp ( )y k kyk
Z g xθ⎛ ⎞= ⎜ ⎟⎝ ⎠
∑ ∑Where
is the normalization factor
is the k-th feature of x
The algorithm IIS‐LLD
Experiments•
Data set: FG‐NET Aging Database–
1002
face images
–
82
subjects–
6‐18
images each
subject–
Age: 0‐69
–
Variations: illumination, pose,
expression, beards, moustaches, spectacles, hats
Experiments•
How to generate LDs from the real age?–
Gaussian
–
Triangle
–
Single
Experiments•
Compared Methods–
AGES, [Geng
et al. TPAMI, 2007]
–
WAS, [Lanitis
et al. TPAMI, 2002]–
AAS, [Lanitis
et al. TSMCB, 2004]
–
General‐purpose classification methods kNN, BP, C4.5, SVM
–
29 college students (human observers)•
HumanA
test
•
HumanB
test
Results•
Mean Absolute Error (MAE)
•
Distribution width
Results•
MAE in different age ranges
Outline•
Introduction
•
Overview of existing techniques•
Evaluation of existing techniques
•
Our approaches•
Conclusions and discussions
Conclusions•
Human age estimation via face is a promising
technique for vast potential applications•
Yet it is quite challenging a task–
Lack of training samples
–
Personalized–
Temporal
•
Comprehensive efforts from both academia and industry have been devoted to algorithm design, modeling, data collecting, system performance test,
valid evaluation protocols, etc.
Conclusions•
Difficulties
(D) vs. Countermeasures
(C)
–
D: Lack of training data–
C: •
Dealing with missing data
[Geng
et al., MM’06, TPAMI’07]•
Learning from label distributions
[Geng
et al., AAAI’10]•
Online training
[Ni et al., MM’09]–
D: Personalized aging patterns
–
C:•
Specified data structure
[Geng
et al., MM’06, TPAMI’07]•
Hierarchical models
[Lanitis
et al., TSMC‐B’04]•
Manifold learning
[Fu et al., TMM’08] [Guo
et al., TIP’08, CVPR’09]
Conclusions•
Difficulties
(D) vs. Countermeasures
(C)
–
D: Temporal aging patterns–
C:•
Specified data structure
[Geng
et al., MM’06, TPAMI’07]•
Regression
[Lanitis
et al. TPAMI’
02, TSMCB’04], [Fu et al., ICME’07, TMM’08], [Yan et al ., ICCV’07], [Zhou et
al., ICCV’05]
Conclusions•
Key components–
Aging face models•
Anthropometric models
•
Active appearance models•
Aging pattern subspace
•
Age manifold•
Appearance feature models
–
Age estimation algorithms•
Classification
•
Regression•
Hybrid methods
•
Learning from label distributions
Conclusions•
Databases–
Publicly available•
FG‐NET Aging Database
•
MORPH Database•
Gallagher’s Web Collected Database
–
Large size•
MORPH Database
•
YGA Database•
LHI Face Database
•
NI’s
Web Collected Database•
Gallagher’s Web Collected Database
Conclusions•
Evaluation protocols–
MAE + Cumulative Scores
–
Leave One Person Out (LOPO)
Future directions•
Data collection–
Extremely laborious, extremely important
–
Ideal data set should •
Cover a wide range of age
•
Include many different subjects•
Contain at least one image for each age of each subject
•
Facial attributes decomposition–
A face image shows multiple facial attributes: identity,
expression, gender, age, race, pose, etc.–
Decomposition of these facial attributes is essential to
extract age‐related features
Future directions•
Multi‐cue age estimation–
People rely on multiple cues to estimate other people’s
age–
Possible indicative cues for age•
Face
•
Voice•
Gait
•
Hair–
Combine face with one or more other cues for age
estimation might remarkably improve the current performance
Future directions•
Pose and illumination variations–
Pose and illumination variations are always troublesome
in real applications–
Pose‐invariant and illumination‐invariant techniques
(intensively investigated in face recognition) might be introduced into age estimation methods
•
Go fuzzy?–
Considering the numerous influential factors, age
estimation can hardly be certain–
Fuzzy classification•
E.g., ‘I am 85% sure that you are 18 years old’
Future directions•
Application oriented age estimation techniques–
Most age‐specific access control systems only need to
determine whether the user is older than a certain age (say 18), which is a binary classification problem
–
‘How old are you?’
‘Are you over 18?’•
Influential factors–
Human aging patterns can be affected by many internal
and external factors, such as genetics, race, gender, health, lifestyle, and even weather conditions
–
Incorporate the influential factors to improve performance
Thank you for attention!
Questions?