Biometric Identification
and
Identical Twins
Kevin W. Bowyer
Computer Science and Engineering
University of Notre Dame
[email protected]
1
CSE X0537 - NOVEMBER 16, 2015.
Twins and Identity Science
Biology
Biometrics
Psychology
How and
to what
extent are
“identical
twins”
different ?
What
biometrics
can tell
“identical
twins” apart,
to what
degree?
How do
people
distinguish
between
“identical
twins” ?
2
CSE X0537 - NOVEMBER 16, 2015.
Distinguishing between pairs of identical twins
is one the more challenging problems in face
recognition.
Klare, Paulino and Jain, Int. Joint Conf. Biometrics 2011.
3
CSE X0537 - NOVEMBER 16, 2015.
https://www.youtube.com/watch?v=XGQAAs5quf4
(classic)
https://www.youtube.com/watch?v=zV-Y0GBL5D8
(modern)
4
CSE X0537 - NOVEMBER 16, 2015.
A detour through a computer
scientist’s understanding of some
biology about twins …
5
CSE X0537 - NOVEMBER 16, 2015.
Frequency of Twins
Ø 
Ø 
Ø 
Twin birth rate: 33.7 per 1,000 live
births in US in 2013 (CDC stats).
About ¾ of twins are “fraternal” and
¼ are “identical”.
More frequent for older mothers, and
with use of fertility drugs; becoming
more frequent overall.
6
CSE X0537 - NOVEMBER 16, 2015.
http://www.empr.com/news/overall-birth-rate-down-but-twin-deliveries-up/article/386846/
CSE X0537 - NOVEMBER 16, 2015.
7
Fraternal Twins
Ø 
Ø 
Ø 
Ø 
Two different fertilized egg cells
(zygotes)
“Dizygotic” or DZ.
No more genetically alike than nontwin siblings. (though they develop in
the womb at the same time)
Share 50% of genes.
8
CSE X0537 - NOVEMBER 16, 2015.
“Identical” Twins
Ø 
Ø 
Ø 
Fertilized egg cell splits into two.
“Monozygotic” or MZ.
Two individuals with same genetic
makeup,
But identical twins are not necessarily
identical in appearance.
9
CSE X0537 - NOVEMBER 16, 2015.
“Mirror” Identical Twins
Ø 
Ø 
Ø 
Single fertilized egg cell splits later
than regular identical (at 9-12 days).
Physical asymmetries expressed
opposite: left / right handed, mirrorimage dental irregularities, …
About ¼ of MZ twins are mirror.
10
CSE X0537 - NOVEMBER 16, 2015.
http://www.twin-pregnancy-and-beyond.com/mirror-twins.html
CSE X0537 - NOVEMBER 16, 2015.
11
Even More Complicated …
Ø 
Ø 
Ø 
Ø 
Identical twins can be monoamniotic
or diamniotic.
Identical twins can be monochorionic
or dichorionic. (placenta)
Monoamniotic is also monochorionic.
Early-splitting zygotes are diamniotic
and dichorionic.
12
CSE X0537 - NOVEMBER 16, 2015.
Point to remember:
Reports of 100% accuracy on
images of a few sets of twins
may be an accident of the
twins being mirror twins.
13
CSE X0537 - NOVEMBER 16, 2015.
A long-held truism of biometrics …
By definition, identical twins cannot be
distinguished based on DNA.
Jain et al, On the similarity of identical twin fingerprints,
Pattern Recognition, 2002.
These twins are genetically identical, with
the same chromosomes … they cannot be
distinguished using the same DNA.
Tao et al, Fingerprint Recognition with Identical Twin
Fingerprints, PLoS ONE, 2012.
14
CSE X0537 - NOVEMBER 16, 2015.
Turns out not to be true given the most
recent DNA analysis techniques …
15
CSE X0537 - NOVEMBER 16, 2015.
Where do you get biometric
datasets of identical twins ?
16
CSE X0537 - NOVEMBER 16, 2015.
www.twinsdays.org
https://www.youtube.com/watch?v=uQJH64QyzPs
CSE X0537 - NOVEMBER 16, 2015.
17
http://www3.nd.edu/~cvrl/CVRL/Data_Sets.html
18
CSE X0537 - NOVEMBER 16, 2015.
How hard is it for face
recognition algorithms
to tell twins apart ?
19
CSE X0537 - NOVEMBER 16, 2015.
Impostor
distribution
for images
of non-twins
Impostor
distribution
for images
of twins
20
CSE X0537 - NOVEMBER 16, 2015.
A to G = 3 top alg’s in MBE 2010 + 4 commercial.
Baseline = Algorithm A with non-twin impostor pairs.
CSE X0537 - NOVEMBER 16, 2015.
21
Best algorithm EER on twins is 13% to 17%.
22
CSE X0537 - NOVEMBER 16, 2015.
Best algorithm EER of 9% to 16%.
23
CSE X0537 - NOVEMBER 16, 2015.
Best algorithm EER of about 7% to 14%.
CSE X0537 - NOVEMBER 16, 2015.
24
Best algorithm EER of about 15% to 16%.
CSE X0537 - NOVEMBER 16, 2015.
25
Point to remember.
Twin discrimination by face
analysis is hard, with room for
improvement that could help
face recognition in general.
26
CSE X0537 - NOVEMBER 16, 2015.
How do humans perform
at distinguishing identical
twins from face
appearance ?
27
CSE X0537 - NOVEMBER 16, 2015.
Twins or images of same person ?
CSE X0537 - NOVEMBER 16, 2015.
28
Twins or images of same person ?
CSE X0537 - NOVEMBER 16, 2015.
29
Twins or images of same person ?
CSE X0537 - NOVEMBER 16, 2015.
30
Twins or images of same person ?
CSE X0537 - NOVEMBER 16, 2015.
31
Twins or images of same person ?
CSE X0537 - NOVEMBER 16, 2015.
32
Humans are more accurate than
current face matching algorithms.
CSE X0537 - NOVEMBER 16, 2015.
33
Humans do better with more time.
CSE X0537 - NOVEMBER 16, 2015.
34
Humans appear to use skin markings
as a major factor in this task.
CSE X0537 - NOVEMBER 16, 2015.
35
Point to remember.
Humans appear to learn to
discriminate twins by
interpreting skin marks in a
“forensic” manner.
(This interpretation is compatible with results by Sarah
Stevenage, British Journal of Psychology, 2011.)
36
CSE X0537 - NOVEMBER 16, 2015.
What about 3D face
analysis for
identical twins ?
37
CSE X0537 - NOVEMBER 16, 2015.
38
CSE X0537 - NOVEMBER 16, 2015.
39
CSE X0537 - NOVEMBER 16, 2015.
Ø 
Ø 
Ø 
Ø 
Ø 
Minolta 910 with “tele” lens.
About 100K points on face.
107 pairs of twins.
Smile + neutral expression.
“Same session” data.
40
CSE X0537 - NOVEMBER 16, 2015.
Alg. 4 = I. Kakadiaris, U. of Houston’s “UR3D”.
Good, but compare to performance on FRGC v2.
CSE X0537 - NOVEMBER 16, 2015.
41
What about irises of
identical twins ?
42
CSE X0537 - NOVEMBER 16, 2015.
“… comparisons among the eyes
of actual monozygotic twins also
yielded a result expected for
unrelated eyes ...”
“How iris recognition works,” John
Daugman, IEEE Trans CVST, 2004.
43
CSE X0537 - NOVEMBER 16, 2015.
Verification of John Daugman’s claim.
CSE X0537 - NOVEMBER 16, 2015.
44
Twins or Unrelated ?
45
CSE X0537 - NOVEMBER 16, 2015.
Twins.
46
CSE X0537 - NOVEMBER 16, 2015.
Twins or Unrelated ?
47
CSE X0537 - NOVEMBER 16, 2015.
Unrelated.
48
CSE X0537 - NOVEMBER 16, 2015.
Twins or Unrelated ?
49
CSE X0537 - NOVEMBER 16, 2015.
Twins.
50
CSE X0537 - NOVEMBER 16, 2015.
Twins or Unrelated ?
51
CSE X0537 - NOVEMBER 16, 2015.
Unrelated.
52
CSE X0537 - NOVEMBER 16, 2015.
Twins or Unrelated ?
53
CSE X0537 - NOVEMBER 16, 2015.
Twins.
54
CSE X0537 - NOVEMBER 16, 2015.
Twins or Unrelated ?
55
CSE X0537 - NOVEMBER 16, 2015.
Unrelated.
56
CSE X0537 - NOVEMBER 16, 2015.
Ø 
76 pairs of twins, plus non-twins.
Ø 
Image pair presented for 3 sec.
Ø 
5-point response scale.
Ø 
Ø 
Over 80% accurate in twins /
non-twin classification.
92% - 93% accuracy for “certain”
responses.
57
CSE X0537 - NOVEMBER 16, 2015.
Point to remember.
Humans readily perceive iris
texture similarity that current
iris recognition technology
does not.
58
CSE X0537 - NOVEMBER 16, 2015.
What about fingerprints of
identical twins ?
59
CSE X0537 - NOVEMBER 16, 2015.
Jain, Prabhakar and Pankanti, On the similarity of identical
twin fingerprints, Pattern Recognition 35, 2653-2663, 2002.
CSE X0537 - NOVEMBER 16, 2015.
60
Essence of Jain et al conclusions:
Ø 
Ø 
Ø 
Twin prints match more closely
than those of unrelated persons.
Twin prints very likely to have
same print category: whorl, …
Twin prints may be like matching
unrelated person prints within
the same category.
61
CSE X0537 - NOVEMBER 16, 2015.
Jain, Prabhakar and Pankanti, On the similarity of identical
twin fingerprints, Pattern Recognition 35, 2653-2663, 2002.
62
CSE X0537 - NOVEMBER 16, 2015.
Point to remember.
Fingerprints do allow reliable
means of distinguishing
between identical twins.
63
CSE X0537 - NOVEMBER 16, 2015.
Twins studies have been done
for various additional modalities:
Ø 
Speaker identification
Ø 
Handwriting
Ø 
Gait
Ø 
Ear
Ø 
…
64
CSE X0537 - NOVEMBER 16, 2015.
Limitations of biometric studies
of twins to date include:
Ø 
Ø 
Ø 
Ø 
“Small” size of dataset
Over-focus on identical twins
without fraternal twins, siblings
Lack of verified MZ status
Focus on one or a small number of
biometrics
65
CSE X0537 - NOVEMBER 16, 2015.
Questions ?
66
CSE X0537 - NOVEMBER 16, 2015.
Figures and tables in this talk are borrowed from:
J. R. Paone, et al, Double Trouble: Differentiating Identical Twins by Face
Recognition, IEEE Trans. Info. Forensics and Security 9 (2), 285-295, February 2014.
http://www3.nd.edu/~kwb/PaoneEtAlTIFS_2014.pdf
S. Biswas et al, A Study of Face Recognition of Identical Twins By Humans, Int’l
Workshop on Information Forensics and Security (WIFS 2011), December 2011.
http://www3.nd.edu/~kwb/BiswasBowyerFlynnWIFS_2011.pdf
V. Vijayan, et al, Twins 3D Face Recognition Challenge, Int. Joint Conference on
Biometrics, October 2011.
http://www3.nd.edu/~kwb/VijayanEtAlIJCB_2011.pdf
P. J. Phillips, et al, Distinguishing Identical Twins By Face Recognition, IEEE Int.
Conf. on Automatic Face and Gesture Recognition (FG 2011), March 2011, 185-192.
http://www3.nd.edu/~kwb/PhillipsEtAlFG_2011.pdf
K.Hollingsworth, et al, Similarity of Iris Texture Identical Twins, IEEE CS Workshop
on Biometrics, June 2010, 22-29.
http://www3.nd.edu/~kwb/HollingsworthEtAlCVIU_2011.pdf
K.W. Bowyer, What Surprises Do Identical Twins Have for Identity Science?, IEEE
Computer 44 (7), July 2011, 100-102.
67
CSE X0537 - NOVEMBER 16, 2015.