Face Age and Demographics

beeuppityΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 4 χρόνια και 21 μέρες)

87 εμφανίσεις

Face Age and Demographics

What’s in a Face?


Age


Gender


Race/ethnicity


Who cares?


These factors may be interesting on their own


They also impact face recognition systems

Impact of Demographics on Face
Recognition


“Face
Recognition Performance: Role of
Demographic Information
,”


B.F.
Klare
, M.J. Burge, J.C.
Klontz
, , R.W.
Vorder

Bruegge
, A.K. Jain, TIFS

vol.7, no.6, pp.1789
-
1801,
Dec. 2012
.


“MBE 2010: Report on the evaluation of 2D still
-
image face recognition algorithms,”


P. J.
Grother
, G. W. Quinn, and P. J. Phillips, National
Institute of Standards and Technology, NISTIR, vol.
7709, 2010.





Demographic Factors

Datasets


Per
-
demographic datasets sampled from
Pinellas County Sherriff's Office (PCSO)
mugshot

database


~8,000 samples for most groups

Algorithms Evaluated


3 commercial off the Shelf (COTS) systems


Black box face recognition systems, no training possible


Training free local feature based methods


LBP


concatenated histograms of LBP values calculated
over patches in a grid, LBP features from 2 radii are
concatenated


Gabor features


concatenated histograms of (quantized)
phase values calculated over patches in a grid


4SF algorithm


Trainable algorithm, using LBP as base features


Gabor Filters





Complex exponential weighted by a Gaussian


Euler’s Theorem:
e
ix

=
cos
(
x
) +
i

sin(
x
)


Output of a Gabor filter is a complex number, can
be viewed as


Orthogonal real part (Gaussian *
cos
) and imaginary
part (Gaussian * sin)


magnitude and phase (polar form)

4SF Algorithm


Spectrally Sampled Structural Subspace
Features (4SF) algorithm

Per
-
Gender True Accept Rates (TAR) at
.1% False Accept Rate (FAR)

Per Race/Ethnicity TAR at .1% FAR

Per Age Group TAR at .1% FAR

Conclusions


Worse performance on female, Black, and
younger cohorts for all algorithms tested


Training per
-
demographic models improved
performance slightly, but did not eliminate
performance differences


MBE 2010: Report on the evaluation of
2D still
-
image face recognition
algorithms



MBE 2010: NIST competition (similar to FRVT)


7 Commercial, 3 academic algorithms
submitted to NIST, evaluated on operational
face recognition databases


Large datasets


~1.8 million law enforcement
book images, ~1.8 million visa images


Various “investigations”

Investigation 15: Link between sex and
accuracy








Marginally worse performance on males

Investigation 16: link between subject
age and accuracy


Investigation 19: Link between race
and accuracy


MBE 2010 Conclusions


Impact of gender:


Generally marginally worse performance on
femalse

than males


Impact of age:


Variable effect, often small not consistent across
algorithms


Impact of race:




Blacks are easier to recognize than
whites
for 5 of
the 6 algorithms. American Indians and Asians were
clearly easier to recognize for 3 of the algorithms
(P01, Z05, and Too), while for V07 American Indians
and Asians appeared more difficult to recognize.


Inconsistencies


MBE 2010 report conclusions on gender, and race are wildly
inconsistent with results reported in the TIFS paper
(although the 3 COTS used in the TIFS paper were also
evaluated in MBE 2010).


Experimental method:
590,105 genuine comparisons from
the FBI set were separated by race. For each algorithm, and
for each race, a distribution of FNMRs (at FMR=0.001) was
computed using 2000 bootstrap iterations. The resulting
plot shows how the false non
-
match rate differs for
different races. A given genuine comparison was
only
retained
if the recorded race was consistent across all
image captures for the given individual. The false match
rate was computed using results from study 3.


Speculation
: The imposter scores used for each group were
drawn from the complete dataset

Within Group vs. Between Group
Imposter Score Distributions


FaceVACS scores on a full comparison matrix of
5,000 probe vs. 5,000 gallery images drawn from
the PCSO dataset


Mean within
-
race imposter score: .0769


Mean cross
-
race imposter score: .0487


Proportion of imposter scores > .5:


Within race: .15 %


Cross race: .0083%


A demographically
-
imbalanced data set has fewer
plausible imposters for minority groups

Age Estimation from Face Images


Given a face image, predict the subjects age


Who cares?


Various possible applications, again including
improving recognition accuracy



AGES


Geng
,
Xin
,
Zhi
-
Hua

Zhou, and Kate Smith
-
Miles
. “Automatic
age estimation based on
facial aging patterns
.”
Pattern Analysis and
Machine Intelligence, IEEE Transactions on
29.12 (2007):
2234
-
2240


FG
-
NET dataset


1,002 images of 82 subjects, avg. 12.2 per
subject





Representation


Given images of a single subject:


Extract features:


AAM PCA shape and appearance coefficients


Apply LDA to the AAM coefficients using sample ages as
class labels (
AGES
lda
)


Form “Aging Pattern,” concatenated features from
every year of the subjects life with some missing
values


Age range 0
-
69, avg. 12.2 samples per subject in
FG
-
NET => aging pattern is mostly missing values


AGES Overview


For an aging patterns
x
k
, find a subspace
projection:

𝒌
=

𝐖
T
(

𝑘



μ
)



x
k

= {

𝑘
𝑎
,

𝑘
𝑚
}


Since
x

is mostly missing values just applying
PCA won’t work


Reconstruction:


𝒌
=
𝜇
+
𝐖

𝑘




𝒌
= {


𝑘
𝑎
,


𝑘
𝑚
}


Goal: Find W which minimizes:




Training Procedure


“E
-
M like algorithm”


Initialization:


Replace

𝑘
𝑚

with [
𝝁
𝑘
𝑚
]


Apply PCA to get
W
0

and
μ
0


Repeat until convergence:


Find
y
k

the least squares solution of:



Calculate


𝒌
, replace

𝑘
𝑚

with


𝑘
𝑚


Apply PCA to get
W
i
+1

and
μ
i




Subspace Based Age Estimation


Given W and
μ

which minimize:



How to predict the age of a test image?


Generate p aging patterns
z
j

(j=1:p), calculate
corresponding
y
j

via:


Select the age which minimizes reconstruction
error from projecting
z
j

into the subspace:



Age Estimation Performance Metrics


Mean Absolute Error (MAE):


Summary statistic:
1
𝑁

𝐴𝑔𝑒
 

𝐴𝑔𝑒
𝑖𝑚𝑎
𝑁
𝑖
=
1


Cumulative Score (CS) Plot:


x
-
axis: absolute error (years)


y
-
axis: fraction of test data with error ≤ x
-
value

Results


FG
-
NET (LOPO)

MORPH Album 1

Human Age Estimation on FG
-
NET
Dataset


This paper uses age estimates from 29
humans on 51 face images, MAE: 6.23


Results from mechanical
turk
, 5 human
estimates per image for all 1,002 FG
-
NET
images: MAE: 4.7

Imbalanced age estimation


FG
-
NET is biased towards young ages


Build per age
-
range AGES models:






4.15 MAE if GT age groups are used, only
slight improvement otherwise


“Age
Synthesis and Estimation via
Faces: A
Survey”


Yun
Fu;
Guodong

Guo
;

Huang, T.S
.; Pattern
Analysis and Machine Intelligence, PAMI
2010


Human facial aging


Age image synthesis


Age estimation



Facial Aging


2 main stages


Early growth


Changes characterized by craniofacial growth


In age estimation, shape based features are good for images
of children


Adult aging


Some shape change over large time periods, but relatively
minor


Skin texture changes


“skin becomes thinner, darker, less
elastic, and more leathery”


Wrinkles form


In age estimation, shape based features less effective for
adults


Age Synthesis


Given an image, generate a synthetic image of
the subject at an older or younger age


Typically need a model:


Geometry
-
Based


ASM, Facial Action Coding System (FACS), caricature
generator


Image based


Clone face attributes from source to target image, Merging
Ratio Images


Appearance
-
Based Models (shape and texture)


AAM


3D
Morphable

model

Age Synthesis Methods


Given a face model, how to generate synthetic
images?


Explicit Data Driven Synthesis


Given a shape model, apply a transform to simulate aging
(PCA based caricature model, growth model


Explicit Mechanical Synthesis


Given a texture model, wrinkle synthesis, structural face
model (
explictly

modeling layers of skin and muscle)


Implicit Statistical Synthesis


Based on Shape + Texture models, variations due to aging
modeled statistically






Face Representations


Anthropometric models

statistics derived from e.g.
ratio of distances between
keypoints
, normalized
keypoint

locations (shape based representation


AAM based features

fitting an AAM gives appearance,
and shape coefficients that together describe a face


Manifold learning


given a set of subjects with known
ages, learn a low
-
dimensional representation capturing
the age distribution


Local feature based methods


LBP based features,
Gabor Features, BIF features (variation of Gabor
features)


Spatially flexible Patch (SFP)


patch location encoded
along with descriptors

Age Estimation Methods


Classification based methods


Consider age prediction an N class classification problem, apply
some classifier


Doesn’t account for the relationship between class labels


Regression based methods


Consider age prediction a regression problem, predict a
numerical value for age


Ages are (typically) discrete values, also constrained to a very
specific numerical range


Ordinal regression techniques can also be applied, e.g.


Kuang
-
Yu Chang; Chu
-
Song Chen; Yi
-
Ping Hung; , "Ordinal
hyperplanes

ranker with cost sensitivities for age
estimation,"

Computer Vision and Pattern Recognition (CVPR),
2011 IEEE Conference on

, vol., no., pp.585
-
592, 20
-
25 June
2011


Hierarchical Methods


Different features may be better for
differentiating coarse age groups than
predicting exact ages


Possible hierarchical models:


Age group classification
-
> within group regression


Rough age prediction via regression
-
>
classification within some range around estimate


Full age range regression
-
> regression within a
restricted range around the initial estimate


Impact of Demographics on Age
Estimation


Guodong

Guo
;
Guowang

Mu; , "Human age estimation: What is the
influence across race and gender?,"

Computer Vision and Pattern
Recognition Workshops (CVPRW), 2010 IEEE Computer Society Conference
on

, vol., no., pp.71
-
78, 13
-
18 June
2010


Experiments on MORPH Album 2 (public
mugshot

database)


Higher age estimation error when training and testing on
different demographic groups


Gender classification
-
> Race Classification
-
> Age Estimation
improved MAE vs. direct age estimation


Guodong

Guo
;
Guowang

Mu; , "Simultaneous dimensionality reduction and human age
estimation via kernel partial least squares regression,"

Computer Vision and Pattern
Recognition (CVPR), 2011 IEEE Conference on

, vol., no., pp.657
-
664, 20
-
25 June
2011


Simultaneous gender, race, and age estimation using KPLS on
BIF features, better age estimation performance than 3 stage
age estimation process