Can Facial Uniqueness be Inferred from Impostor Scores? · Can Facial Uniqueness be Inferred from...
Transcript of Can Facial Uniqueness be Inferred from Impostor Scores? · Can Facial Uniqueness be Inferred from...
Can Facial Uniqueness be Inferredfrom Impostor Scores?
Abhishek Dutta
Oct 14, 2013, Nijmegen, Netherlands.
Presentation accompanying the following paper:A. Dutta, R. Veldhuis and L. Spreeuwers. Can Facial Uniqueness be Inferred from Impostor Scores?, BTFS 2013,Nijmegen, Netherlands.
A quick question
Image Pair 1
Are they same subject ?
A quick question ...
Image Pair 2
What about this image pair?Are they same subject?
Facial Uniqueness (or, Distinctness) in a Population
Image Pair 1 Image Pair 2
person-id261
person-id263
person-id272
Irrespective of your decision {yes, no, maybe} , you were:I more uncertain in your decision for Image Pair 1,I less uncertain in your decision for Image Pair 2.
Why ?
Facial Uniqueness (or, Distinctness) in a Population ...
Image Pair 1 Image Pair 2
person-id261
person-id263
person-id272
If we assume that faces form a Gaussian cloud in a highdimensional space, then
I identity relates to direction, andI distinctness relates to distance from the mean.
Therefore, we are more confident in making decision about identitywhen facial images are well separated in the face space – by virtueof their distinctness (as in Image pair 2).
Facial Uniqueness (or, Distinctness) in a Population ...
I Therefore, in Biometrics, researchers have been activelyinvolved in measuring uniqueness from facial photographs
I Such facial uniqueness measurements are useful to build anface recognition system that adaptively vary the decisionthreshold to improve recognition performance.
Measuring Facial Uniqueness
Impostor Populationhow unique is
this facialimage? with respect to this
impostor population?
similarity score
coun
t
coun
t
similarity score
fairly commonface
highly distinctface
Measuring Facial Uniqueness ...
Impostor Populationhow unique is
this facialimage? with respect to this
impostor population?
similarity scoreco
unt
coun
t
similarity score
fairly commonface
highly distinctface
AssumptionI similarity score is influenced only by facial identity
I non-unique facial image1 will generally exhibit high level ofsimilarity to many other subjects in a large population (bydefinition)
Hypothesis facial uniqueness of a subject can be inferred from itsimpostor similarity score distribution
1lamb in the biometric zoo
Can Facial Uniqueness be Inferred from Impostor Scores?
In this paper, we investigate the stability of facial uniquenessmeasures based on impostor scores. For this, we perform followingtwo experiments:
Experiment 1
To study the influence of image quality variations on the impostorscore distribution.
Experiment 2
To study the stability of a recently proposed Impostor BasedUniqueness Measure (IUM) of [Klare and Jain 2013] under imagequality variations.
Exp. 1: Influence of Image Quality on Impostor Scores
Probe(Average Face)
Baseline Gallery
Motion Blur(angle = 0)
len. = 09 len. = 31
Gaussian Noise(mean= 0)
var. = 0.07 var. = 0.3
Pose(camera-id)
05_1 04_1 19_014_013_008_0 05_0
19_108_1
Gallery with Image Quality Variation
Baseline ImpostorScore Distribution
Quality based ImpostorScore Distribution
Exp. 1: Results
Pose Blur (Motion) Noise (Gaussian)
0.0
0.1
0.2
0.3
0.4
0
20
40
60
80
−0.1
0.0
0.1
0.2
−2
−1
0
1
2
3
FaceV
AC
SV
erilo
ok
LR
PC
AcLD
A
08
_1
08
_0
13
_0
14
_0
05
_1
05
_0
04
_1
19
_0
19
_1 0 3 5 7 13
17
29
31
0
0.0
07
0.0
3
0.0
7
0.1
0.3
Quality Variation [ Pose: camera−id | Blur: blur length, angle = 0 | Noise: variance, mean = 0 ]
Sim
ilari
ty s
co
re w
ith
ave
rag
e fa
ce
Pose Blur (Motion) Noise (Gaussian)
0.0
0.1
0.2
0.3
0.4
0
20
40
60
80
−0.1
0.0
0.1
0.2
−2
−1
0
1
2
3
FaceV
AC
SV
erilo
ok
LR
PC
AcLD
A
08
_1
08
_0
13
_0
14
_0
05
_1
05
_0
04
_1
19
_0
19
_1 0 3 5 7 13
17
29
31
0
0.0
07
0.0
3
0.0
7
0.1
0.3
Quality Variation [ Pose: camera−id | Blur: blur length, angle = 0 | Noise: variance, mean = 0 ]
Sim
ilari
ty s
co
re w
ith
ave
rag
e fa
ce
I the nature of impostor score distribution corresponding to allthree types of quality variations is significantly different fromthe baseline impostor distribution.
I the impostor score distribution also seem to be responding toquality variations.
Exp. 1: Results
Pose Blur (Motion) Noise (Gaussian)
0.0
0.1
0.2
0.3
0.4
0
20
40
60
80
−0.1
0.0
0.1
0.2
−2
−1
0
1
2
3
Fa
ce
VA
CS
Ve
riloo
kL
RP
CA
cL
DA
08
_1
08
_0
13
_0
14
_0
05
_1
05
_0
04
_1
19
_0
19
_1 0 3 5 7 13
17
29
31
0
0.0
07
0.0
3
0.0
7
0.1
0.3
Quality Variation [ Pose: camera−id | Blur: blur length, angle = 0 | Noise: variance, mean = 0 ]
Sim
ilari
ty s
core
with a
vera
ge face
I the impostor score distribution of the four systems respond ina different way to the three types of image quality variations.
Exp. 1: Conclusion
Impostor score distribution is not only influenced by identity(as expected) but also by the image quality like pose, blur andnoise.
Exp. 2: Stability of [Klare and Jain 2012] IUM
In this experiment, we study the stability of a recently proposedImpostor Based Uniqueness Measure (IUM) of [Klare and Jain2013] under image quality variations.
SmaxSmin
Smean
similarity score
count impostor score
distribution
IUM score =Smax − Smean
Smax − Smin
Uniqueness2 IUM score
high ∼ 1.0low ∼ 0.0
2with respect to the impostor population
Exp. 2: Stability of [Klare and Jain 2012] IUM ...
query imagefrom session 4
remaining 197 subjectsfrom session 3
remaining 197 subjectsfrom session 4
query imagefrom session 3
1006 subjects fromFERET Fa subset 1039 subjects from
CAS-PEAL pose PM+00
Impostor Population for Session 3 image
Impostor Population for Session 4 image
I We vary the quality of session 4 images (pose, noise, blur)
I If the IUM scores are stable with image quality variations, theIUM scores computed from session 3 and 4 should remainhighly correlated despite quality variation in session 4 images.
Exp. 2: Results
Pose Blur (Motion) Noise (Gaussian)
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
Fa
ce
VA
CS
Ve
riloo
kL
RP
CA
cL
DA
08_1
08_0
13_0
14_0
05_1
05_0
04_1
19_0
19_1 0 5 9 17
31
0
0.0
3
0.0
7
0.1
0.3
Quality Variation [ Pose: camera−id | Blur: blur length, angle = 0 | Noise: variance, mean = 0 ]
Norm
aliz
ed c
orr
ela
tion c
oeffic
ient
Imposter based Uniqueness Measure of [Klare and Jain 2012]is not stable under image quality variations.
Conclusion
I We have shown that impostor score is influenced by bothidentity and quality of facial images.
I We have also shown that any attempt to measurecharacteristics of facial identity (like facial uniqueness) solelyfrom impostor score distribution shape may give misleadingresults in the presence of image quality degradation in theinput facial images.
Future Work
I This research has thrown up many questions in need of furtherinvestigation regarding the stability of existing facialuniqueness measures based solely on impostor scores.
I More research is needed to better understand the impact ofimage quality on the impostor score distribution. Such studieswill help develop uniqueness measures that are robust toquality variations.
ReferencesI Brendan F. Klare and Anil K. Jain, Face recognition: Impostor-based
measures of uniqueness and quality, in Biometrics: Theory, Applicationsand Systems (BTAS), 2012 IEEE Fifth International Conference on, 2012,pp.237244.
I George Doddington, Walter Liggett, Alvin Martin, Mark Przybocki, andDouglas Reynolds, Sheep, goats, lambs and wolves: A statistical analysisof speaker performance in the nist 1998 speaker recognition evaluation, inProceedings of International Conference on Spoken Language Processing,1998.
I Cognitec Systems, FaceVACS C++ SDK Version 8.4.0, 2010.
I Neurotechnology, 2011. VeriLook C++ SDK Version 5.1,
I CSU Baseline Algorithms - Jan. 2012 Releases,http://www.cs.colostate.edu/facerec/ algorithms/baselines2011.php.
Questions and Feedback
http://abhishekdutta.org