1536

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

Analysis of Facial Marks to Distinguish Between
Identical Twins
Nisha Srinivas, Gaurav Aggarwal, Patrick J. Flynn, Fellow, IEEE, and Richard W. Vorder Bruegge

Abstract—Identical twin face recognition is a challenging task
due to the existence of a high degree of correlation in overall facial appearance. Commercial face recognition systems exhibit poor
performance in differentiating between identical twins under practical conditions. In this paper, we study the usability of facial marks
as biometric signatures to distinguish between identical twins. We
propose a multiscale automatic facial mark detector based on a
gradient-based operator known as the fast radial symmetry transform. The transform detects bright or dark regions with high radial symmetry at different scales. Next, the detections are tracked
across scales to determine the prominence of facial marks. Extensive experiments are performed both on manually annotated and
on automatically detected facial marks to evaluate the usefulness of
facial marks as biometric signatures. Experiment results are based
on identical twin images acquired at the 2009 Twins Days Festival
in Twinsburg, Ohio. The results of our analysis signify the usefulness of the distribution of facial marks as a biometric signature.
In addition, our results indicate the existence of some degree of
correlation between geometric distribution of facial marks across
identical twins.
Index Terms—Face recognition, facial marks, identical twins.

I. INTRODUCTION

T

HE ability to distinguish between identical twins based
on different biometric modalities such as face, iris, fingerprint, etc., is a challenging and interesting problem in the
biometric area [1]–[5]. Identical twins (also known as monozygotic twins) are formed when a zygote splits and forms two embryos. They cannot be discriminated based on DNA. Therefore,
other biometric traits are needed to distinguish between identical twins. Using face recognition to differentiate between identical twins (monozygotic twins) is very difficult [3], because of
the high degree of similarity in their overall facial appearance.
In this paper we focus on distinguishing between monozygotic
twins based on localized facial features known as facial marks.
Traditionally, biometrics research has focused primarily on
developing robust characterizations and systems to deal with
challenges posed by variations in acquisition conditions (such
Manuscript received December 08, 2011; revised June 04, 2012; accepted
June 11, 2012. Date of publication June 29, 2012; date of current version
September 07, 2012. This work was supported by the U.S. Department of
Justice/National Institute of Justice under Grant 2009-DN-BX-K231. The
associate editor coordinating the review of this manuscript and approving it for
publication was Dr. Jaihie Kim.
N. Srinivas, G. Aggarwal, and P. J. Flynn are with the Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN
46556 USA (e-mail: nsriniva@nd.edu; gaggarwa@nd.edu; flynn@nd.edu).
R. W. Vorder Bruegge is with the Federal Bureau of Investigation, Digital
Evidence Laboratory, Quantico, VA 22135 USA (e-mail: Richard.VorderBruegge@ic.fbi.gov).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIFS.2012.2206027

Fig. 1. A pair of identical twins from the identical twins dataset. We observe
a high degree of overall facial similarity and the difference in the number and
type of facial marks.

as pose, illumination condition, distance from sensor, etc.) and
the presence of noise in the acquired data [6]. Only recently
have researchers started to look at the challenges involved in
dealing with the task of distinguishing between identical twins
[1], [3]. Developing techniques and systems that improve twin
face recognition should also improve generic face recognition
systems. Although identical twins represent only 0.5% of the
global population [2], failure to correctly identify each twin has
led to problems for law enforcement agencies [1]. There have
been several criminal cases in which either both or neither of
the identical twins was convicted due to the difficulty in determining the correct identity of the perpetrator [1].
In this paper, we propose to differentiate between identical
twins using facial marks alone. Facial marks are considered to
be unique and inherent characteristics of an individual. Fig. 1
shows a pair of identical twins from the dataset. Although they
are similar in appearance, they can be distinguished using facial
marks. High-resolution images enable us to capture these finer
details on the face [7]. Facial marks are defined as visible changes
in the skin and they differ in texture, shape and color from the
surrounding skin [8]. Facial marks appear at random positions
of the face. By extracting different facial mark features we aim
to differentiate between identical twins. We have defined eleven
types of facial marks including moles, freckles, freckle groups,
darkened skin, lightened skin, etc., for the analysis.

1556-6013/$31.00 © 2012 IEEE

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

1537

Fig. 2. Approach to distinguish between identical twins using manually detected facial marks.

Fig. 3. Overview of the proposed multiscale automatic facial mark detection process.

Initially, each image in the identical twin dataset is manually
annotated by multiple observers to determine the different types
of perceptible facial marks. Manually annotated facial marks are
characterized both by location and category. The approach to distinguish between monozygotic twins based on manually detected
facial marks is shown in Fig. 2. Next, we propose a multiscale
automatic facial mark detector based on the fast radial symmetry
transform (FRST) [9]. The transform detects dark regions with
high radial symmetry. An overview of the proposed multiscale
automatic facial mark detector is shown in Fig. 3. Initially, an
image is represented at multiple scales in the form of a Gaussian
pyramid. An Active Shape Model (ASM) [10] is used to detect the
contours of the primary facial features like eyes, lips, nostrils and
eyebrows. Using the output of the ASM, a mask is created to remove the primary facial features. Next, the FRST is applied to the
masked image to detect dark regions with radial symmetry. The
aforementioned steps are applied to all images in the Gaussian
pyramid. Finally, the detections are tracked across scales. Automatically detected facial marks are characterized only by geometric location. The locations of facial marks are converted from
image pixel coordinates to barycentric coordinates [11] to facilitate interimage comparison of landmark locations. The similarity
in the distribution of facial marks is used to determine the similarity between two face images. The similarity is computed by
formulating a bipartite graph matching problem.

Extensive experiments are conducted on the manually and
automatically detected facial marks. The data used for the investigation was acquired at the Twins Days Festival at Twinsburg, Ohio in 2009 [3]. The dataset consists of 477 images corresponding to 178 subjects and 89 pairs of twins. The presented
results indicate the need for an automatic facial mark detector
and demonstrate that facial marks can be used biometric signatures to distinguish between identical twins. Prior research has
claimed that the number of facial marks between twins is similar
but the distribution of facial marks across twins is different [12].
We also analyze this conjecture. Contrary to the commonly held
belief, our results indicate nontrivial correlation between distributions of facial marks across identical twins.
A preliminary version of this investigation was presented in
[13]. In [13], we presented the use of facial marks as biometric
signatures by analyzing facial marks annotated by multiple observers. The dataset used in [13] consisted of only 295 twin face
images and 76 pairs of twins. In this paper we have introduced
a multiscale automatic facial mark detector based on the fast
radial symmetry transform and evaluated its performance. We
also compare performance across multiple annotation sessions
and multiple observers. The different experiments are executed
on a larger dataset.
The paper is organized as follows. Section II discusses related
work. Section III provides description of different categories

 1538

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

of facial marks. Section IV describes the identical twin dataset
used in this investigation. The manual annotation process is described in Sections V and VII describes the proposed multiscale automatic facial mark detector. Section VI describes the
matching process. The details of the experimental setup and results are presented in Section VIII. The paper concludes with a
brief summary and discussion.
II. RELATED WORK
We first discuss prior research related to facial mark detections and then focus on identical twin face recognition research.
A. Related Work on Facial Mark Detections
Lin et al. [7] represented the face at multiple layers in
terms of global appearance, facial features, skin texture and
irregularities that contribute towards identification. Global
appearance and facial features are modeled using a multilevel
PCA (Principal component analysis) followed by regularized
LDA (Linear discriminant analysis). A Scale Invariant Feature
Transform (SIFT) is employed to detect and describe details
of irregular skin region, which is combined with elastic graph
matching for recognition. Improved performance was achieved
by fusing facial features at multiple levels. Pierrard et al.
[14] presented a framework to localize prominent facial skin
irregularities, like moles and birthmarks. They use a multiscale template matching algorithm for face recognition. A
discriminative factor is computed for each point by using skin
segmentation and local saliency measure and is used to filter
points.
Recently, Zhang et al. [15] designed a facial skin mark
matcher based on a region growing algorithm. Each facial
mark is described in terms of position, color intensity and size.
The results of the facial skin mark matcher are fused with the
results of a PCA based matcher to evaluate the performance of
the system. A drawback of the method is that they need to have
prior knowledge about the location of facial marks in order to
apply the region growing algorithm. Park et al. [16] proposed
to use facial marks as soft biometrics. They initially map each
face image contour obtained from an Active Appearance Model
(AAM) into a mean shape using barycentric texture mapping.
The mean shape images are then filtered using a Laplacian of
Gaussian filter. This acts as a blob detector. Once the facial
marks are detected, matching is performed based on Euclidean
distance with a set threshold. The number of matches represents
the similarity score between images. However, they do not
use facial marks by themselves to evaluate performance. They
fuse the scores from the facial mark matcher and a commercial
face recognition software to evaluate the performance. They
observed marginal improvement in performance by fusing the
two scores.
We have not compared the proposed method with any other
previously published methods mainly because previous approaches fuse different face features with the features obtained
from facial marks. The work proposed in this paper uses only
facial marks as biometric signatures to distinguish between individuals. Most previously published work use low-resolution
images. The number of facial marks detected in low-resolution

images is generally lower than the number of facial marks
detected in high-resolution images. Since a higher number of
facial marks are detected using high-resolution images, we are
able to represent an individual by a larger and more unique
feature set. Therefore, the performance of the low-resolution
approaches is not directly comparable to the proposed approach.
B. Related Work on Identical Twin Biometrics
Recently researchers have started to look at the challenges involved in dealing with the task of distinguishing between identical twins. Kong et al. [17] observed that palm prints from identical twins have correlated features (though they were able to
distinguish between them based on other nongenetic information). The same observation was made by Jain et al. [4] for fingerprints. They observed that though fingerprints appear to be
more similar for identical twins than unrelated persons, fingerprint matching systems can distinguish between them. Genetically identical irises were compared by Daugman and Downing
[5] and were found to be as uncorrelated as the patterns of irises
from unrelated persons. Kodate et al. [18] experimented with
ten sets of identical twins using a 2-D face recognition system.
Recently, Sun et al. [1] presented a study of distinctiveness
of biometric characteristics in identical twins using fingerprint,
face and iris biometrics. They observed that though iris and
fingerprints show little to no degradation in performance when
dealing with identical twins, face matchers experienced problems in distinguishing between identical twins. All of these
studies were either conducted on very small twin biometric
datasets or evaluated using existing inhouse or commercial
matchers. Phillips et al. [3] presented the first detailed study
on discrimination of identical twins using different face recognition algorithms. They compared three different commercial
face recognition algorithms on the identical twins dataset
acquired at Twins Day festival in Twinsburg, Ohio. The dataset
consists of images acquired under varying conditions such as
facial pose, illumination, facial expression, etc. They observed
that it is easier to distinguish between identical twins under
controlled studio-like settings than under uncontrolled settings.
III. TYPES OF FACIAL MARKS
A facial mark is defined as a region of skin or superficial
growth that does not resemble the skin in the surrounding area.
Facial marks represent finer details on the face. They contain information useful to discriminate between identical twins. Availability of high resolution images enables us to view facial marks
in greater detail for analysis. We have identified and defined the
following facial marks (shown in Fig. 4),
1) Mole: A small flat spot less than 1 cm in diameter. The
color of a mole is not the same as the nearby skin. It appears
in a variety of shapes and is normally black in color.
2) Freckle: A small flat spot less than 1 cm in diameter and
appears in a variety of shapes. It is usually brown in color.
3) Freckle group: A cluster of freckles.
4) Lightened patch: A flat spot that is more than 1 cm in
diameter and appears in different shapes. It is lighter in
color than its surroundings.

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

Fig. 4. The different categories of facial marks defined.

5) Darkened patch: A flat spot that is more than 1 cm in
diameter and appears in different shapes. These spots are
darker in color than their surroundings.
6) Birthmark: A persistent visible mark on the skin that is
evident at birth or shortly thereafter. Birthmarks are generally pink, red, or brown in color.
7) Splotchiness: An irregularly shaped spot, stain, or colored
or discolored area.
8) Raised skin: A solid, raised mark less than 1 cm across.
It has a rough texture and appears red, pink, or brown in
color.
9) Scar: Discolored tissue that permanently replaces normal
skin after destruction of the epidermis.
10) Pockmark: A hollow area or small indentation.
11) Pimple: A raised lesion that is temporary in nature.
IV. DATA
The dataset consists of face images of identical twins acquired in two days of August, 2009 at the Twins Days Festival in Twinsburg, Ohio [3]. Face images were captured under
different scenarios and conditions like controlled and uncontrolled lighting, presence and absence of eyeglasses, different
facial expressions like smile or neutral, different poses with yaw
ranging from
to 90 degrees, where 0 degrees is a frontal
view. The dataset used for the proposed experiments consists
of only frontal
face images with no glasses, no facial hair and a neutral expression. These images were captured
under controlled lighting. The 2009 dataset consists of 477 images corresponding to 178 subjects and 89 pairs of twins. Fig. 1
shows images of a set of identical twins from the dataset. The
guidelines used to capture facial images at the Twins Day Festival match the requirements defined by SAP level 51 [19]. The
resolution of the images is 4310 2868
. The average
interpupillary distance is 567 pixels. These high-resolution images enable us to observe finer details on the face when compared to low-resolution images.
V. MANUAL ANNOTATIONS
To evaluate the usefulness of the facial marks, the twins
dataset was annotated initially by multiple observers. Manual
annotations enable us to gain insight on the different categories
and number of facial marks present in the dataset, the location

1539

of facial marks and an understanding of how human observers
visualize and differentiate between facial marks. The manual
annotation process is accomplished using MarkIt, a facial
annotation tool developed at our laboratory by Matthew Pruitt.
MarkIt was designed to aid users to manually annotate images,
and has three main components.
1) Display component: The images are displayed so that the
user may observe and annotate the various facial marks.
2) Annotation component: Contains a list of predefined facial
marks for annotations.
3) Tools component: Presents different shapes of bounding
boxes to perform the actual annotations.
The metadata produced by MarkIt for each image consists of
the number, types and locations of annotated facial marks. We
conducted two sessions of manual annotations with a time lapse
of six months. In each session, a different set of images of the
identical twins dataset was annotated. In the first session 275
images were annotated and in the second session 202 images
were annotated for a total of 477 annotated images.
Four observers, denoted 1 through 4, annotated the images in
two sessions. Observers 1, 2 and 3 did the first session and 1,
2, and 4 did the second session. The observers had no prior experience with facial mark annotation. Hence the observers were
provided with the definitions of the facial marks characterized in
this investigation along with example markings on a few sample
images. Fig. 7 shows examples of different facial marks manually annotated by observer 1, observer 2 and observer 3. We
observe that there is a difference in number of facial marks annotated and categorized between observers. Figs. 5 and 6 show
the category-wise distribution of the number of facial marks annotated by each observer in session 1 and session 2. Table I indicates the total number of facial marks annotated by observers in
session 1 and session 2. Based on these gross statistics, there is
large variation in how the four observers perceived facial marks
in the dataset.
Difficulties in manual annotations noted by observers and us
include:
1) Manual annotation is a difficult task because it involves
training and familiarizing an individual with the definitions
and characteristics of the different types of facial marks
(apart from learning how to use the facial annotation tool).
2) Observers experienced difficulty in differentiating between
categories of facial marks, especially in the case of moles
and freckles.
3) Although all observers appear to have annotated the prominent facial marks, some observers failed to annotate the
less prominent but visible marks.
4) It is a time consuming and an expensive task.
Hence there is a need for an automatic facial mark detector
to overcome these problems. Such a method is discussed in
Section VII.
VI. FACIAL MARKS BASED MATCHING
We propose a matching approach that characterizes each face
image based on the corresponding mark locations between the
gallery and probe images. Each facial mark is characterized by a
facial mark category and geometric location on the face image.
Mark locations are transformed from image space to normalized

 1540

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

avoid correspondence between facial marks of different categories, edges connecting marks belonging to different categories are given infinite weight. A potential match is determined
if the Euclidean distance between a pair of feature centroids (belonging to same category) is less than a threshold . The optimal correspondences are then established by executing standard Hungarian bipartite graph matching algorithm [21] on the
set of potential matches. The normalized similarity score
for
the comparison between image and image is given by
(1)
Fig. 5. Variation in number of facial marks of different types identified by each
observer during session 1.

where
is the total number of correspondences established
across the two images being compared. The similarity metric
is used as a match score between two images.
B. Preliminary Results: Manual Annotations

Fig. 6. Variation in number of facial marks of different types identified by each
observer during session 2.

barycentric coordinate space. Barycentric coordinates are normalized homogeneous coordinates that are scale, rotation and
translation invariant. They are computed with respect to a reference triangle that forms the basis for the transformation. In
this work, the reference triangle is determined by the center of
each eye and the tip of the nose, which is automatically localized
using STASM [20]. Given the reference triangle, points inside
and outside the reference triangle can be expressed as a function
of the vertices of the triangle. Since barycentric coordinates are
computed using three vertices, it provides robustness to facial
pitch changes.
A. Bipartite Graph Matching
Similarity between two sets of facial marks is computed by
representing the matching problem using a weighted bipartite
graph. A weighted bipartite graph is defined as
, where and denote two disjoint sets of vertices and
denotes the connecting edges with corresponding nonnegative
cost. A bipartite graph with
nodes corresponding to
facial marks in image
and
facial marks in image
is constructed. The edges correspond to potential matches between facial marks in and . The nonnegative weight associated with each edge is a function of Euclidean distance between
normalized feature locations being compared. Facial marks of
the same category should be compared against each other (e.g.,
mole against mole, freckle against freckle, etc.). Therefore, to

We perform extensive experiments on the manually annotated
dataset. Each experiment is designed and implemented such that
it evaluates the usefulness of facial marks as a biometric signature to distinguish between identical twins. The outcome of each
of the experiments provide answers to the following questions,
1) How do facial marks annotated by a single observer distinguish between identical twins?
2) Is performance consistent when comparing annotations
from different observers against each other?
3) Is performance consistent when comparing annotations
from the same observer in different sessions?
The different experiments are designed as follows:
1) Experiment 1
a) Experiment Setup: Compare the query set against the
target set annotated by the same observer in the same
session.
b) Target Set: Session Data, where represents the
session number and
or 2.
c) Query Set: Session Data
d) Inference: Facial marks can be used as potential
biometric signatures to distinguish between identical
twins.
2) Experiment 2
a) Experiment Setup: Compare the query set against the
target set annotated by the same observer in different
sessions
b) Target Set: Session 1 Data
c) Query Set: Session 2 Data
d) Inference: Performance degrades due to inconsistency in annotations made by an observer across
sessions.
3) Experiment 3
a) Experiment Setup: Compare the query set annotated
by one observer against the target set annotated by another observer. For example, a query set annotated by
observer 1 is compared against a target set annotated
by observer 2, and the process is known as observer
1 versus observer 2.
b) Target Set: Session Data, where represents the
session number and
or 2.
c) Query Set: Session Data

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

1541

Fig. 7. Different facial marks manually annotated by each observer in session 1 for a given image. (a) observer 1; (b) observer 2; (c) observer 3.

TABLE I
TOTAL NUMBER OF FACIAL MARKS ANNOTATED BY EACH OBSERVER DURING
SESSION 1 AND SESSION 2 OF MANUAL ANNOTATIONS

d) Inference: Performance degrades due to inconsistency in perceiving facial marks by different
annotators.
Fig. 8 compares the Receiver Operating Characteristic
(ROC) curves for Experiments 1, 2, and 3. We observe that
the best performance is obtained when comparing annotations
made by a single observer in the same session, i.e., Experiment
1. This indicates that facial marks are useful in distinguishing
between identical twins. However, performance curves obtained for Experiments 2 and 3 exhibit a significant degradation
in performance, indicating that individual observers perceive
facial marks differently over time and the annotations are
inconsistent. Similarly, different observers view facial marks
differently, leading to inconsistency. Also, the annotation style
of an observer varies differently over time. Inconsistency in
performance observed in Experiment 2 and Experiment 3 is a
major drawback of using manually annotated facial marks to
differentiate between identical twins. Hence, in order to obtain
consistency, there is a need for a robust and efficient automatic
facial mark detector. In the following sections we present such
a multiscale automatic facial mark detector.
VII. AUTOMATIC FACIAL MARK DETECTION
Fig. 9 presents a detailed overview of the proposed automatic facial mark detector. Facial mark detection is performed
at different scales, which is achieved by constructing a Gaussian
pyramid [22]. For each image in the pyramid the following steps
are applied:

Fig. 8. Performance curves obtained for the different types of experiments executed on the manually annotated dataset.

1) The primary facial features like eyes, eyebrows, lips and
nostrils are localized using an Active Shape Model (ASM).
2) Individual masks are created based on the output of the
ASM to mask the primary features.
3) A gradient-based interest operator (the fast radial symmetry transform) is applied to the masked image. The
transform detects regions of high radial symmetry.
4) Applying a threshold to the output of the fast radial symmetry transform results in detecting bright or dark regions
of high radial symmetry, which corresponds to potential facial marks.
The potential facial marks are detected at each level of the image
pyramid. Those present in two or more levels are considered
to be reliable facial marks. Presently, we do not categorize the
detections into different types; we just treat them as point features. In the following sections we describe the process of image
pyramid construction, localization of primary facial features,
mask generation, and the fast radial symmetry transform.

 1542

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

Fig. 9. Diagram highlighting the main steps of the proposed multiscale automatic facial marks detection process.

A. Gaussian Pyramid Construction
The objective is to detect facial marks that are stable across
different scales. This can be achieved by using a Gaussian
pyramid. The Gaussian pyramid consists of a set of low-pass
filtered and subsampled images. The original image is defined
at the base level. The successive levels of the pyramid are
obtained by filtering the image in the previous level and downsampling it by a factor 2. Gaussian pyramid [22] is defined by

(2)

(3)
where
is the base image of size
and
represents the images in the subsequent levels, is the level
number and
is a Gaussian filter of size 5 5.
The number of levels in a Gaussian pyramid is defined by
. In this study we define
. Facial marks are
detected at each level and then tracked across levels to signify
their prominence.
B. Detection of Primary Facial Features
The contours of primary facial features like eyes, eyebrows,
nostrils and lips are detected using an Active Shape Model.
Primary facial features must be masked before the detection
process to avoid detections that are caused by their presence.
The Active Shape Model was first presented by Cootes et al.
[10]. Once an ASM is trained based on the data found in the
training set, it iteratively deforms a contour to fit the new image.

The ASM defines two components of an object, the model
shape and the profile. Model shape defines the shape of the contour. The profile is defined for each contour point and describes
what the image looks like around each point in the model. We
use an open source implementation of ASM called STASM
[20]. This detects 68 facial landmark points corresponding
to the contours of the primary facial features. Using these
landmark points, a mask is created for each image to mask out
the primary facial features referred to as masked images.
C. Fast Radial Symmetry Detector
We propose to apply the fast radial symmetry transform, to
detect the desired facial marks. FRST was defined by Loy et al.
[9]; it is similar to a generalized symmetry transform [23] and
the circular Hough transform [24]. The output of the transform
highlights radially symmetrical regions and suppresses regions
that are asymmetrical. For a given image , FRST determines
the contribution of each pixel to the symmetry over a set of
radii
. The radii set
is defined based on the size of
the different facial marks that appear on the face. Also, at each
image level a different set is defined.
Fig. 9 illustrates the multiscale automatic detection process.
Initially, images are represented at different scales. Next for
each image, the gradient image, is computed using a 3 3
Sobel operator. We compute the orientation projection image
and the magnitude projection image
at each point of
the gradient image for every value in . These images are determined by observing the gradient at each point from which
a corresponding positively affected pixels and negatively-affected pixels are determined. Positively affected pixels are defined as pixels that lie along the direction of the gradient vector
and negatively-affected pixels are pixels in the direction directly opposite to the gradient vector as shown in Fig. 10. The

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

1543

coordinates of the positively-affected and negatively affected
pixels are defined as
(4)
(5)
where
is the gradient vector and
is the magnitude of the gradient vector. Initially the orientation projection
image
and the magnitude projection image
are zero.
The points in
and
corresponding to a pair of positively-affected or negatively-affected pixels is computed by
(6)

Fig. 10. Positively and negatively affected pixels influenced by point
.
gradient image for

of the

(7)
(8)
(9)
The orientation projection image and magnitude projection
image capture the radial features of a particular region. The
radial symmetry contribution , at radius is given by
(10)
where

composition (SVD) of the gradient matrix
region gradient matrix is defined as [26]:
..
.

..
.

..
.

..
.

of the region. The

(14)

and the local dominant orientation information is obtained by
computing the SVD of ,

is given by

(15)

(11)
is
,

(12)

where
is the Gaussian kernel, is the degree of radial strictness and is the normalizing factor. The final symmetry image
is formed by averaging the radial symmetry distributions over
all radii
as
(13)
After the radial symmetry image is computed, we apply hysteresis thresholding to detect dark or bright regions of high radial symmetry resulting in a binary image
.
Then we apply connected component analysis to detect connected regions which correspond to potential facial marks. For
each image, we carry out the above procedure at different levels
of the pyramid, hence we have a set of potential facial marks
at each level. Next, we track the potential facial marks across
levels to form the final set of detections. Detections present at
two or more levels are included in the final detection set. Once
the final detection set is calculated, we perform postprocessing
to reduce the number of false positives.
False positives occur mainly due to presence of hair on the
forehead. These are eliminated by using dominant orientation
information obtained from local gradients of the detected regions [25]. The dominant orientation information for each detected region is computed by calculating the Singular Value De-

where is an orthonormal to . The singular values and
describe the energy in the direction of vectors
and . If the
energy in the direction of vector
for a given region,
then we consider that region as a hair region and is eliminated.
Also, if the size of the detections are greater than a predefined
value
then the detections are eliminated. The remainder of
the detections are characterized as facial marks. In Fig. 11, the
detections represented by the square markers represent the potential facial marks detected (before postprocessing). The detections represented by the circular marker are eliminated using
SVD and the detections represented by the square marker indicate the true positives. The different colors indicate the number
of scales at which the facial marks were detected. Facial marks
detected are shown in Fig. 12. Currently, each facial mark is defined based only on the geometric location and we do not classify them into different categories. They are treated as point features. Once the point features are computed we perform facial
mark matching.
D. Bipartite Graph Matching for Automatically Detected
Facial Marks
The process for matching facial marks detected by the multiscale automatic facial mark detector is similar to that described
in Section VI. In the case of automatically detected facial marks,
each facial mark is characterized only by its geometric location on the corresponding face image. Therefore, automatically
detected facial marks are treated as point features and can be
viewed as they all belong to the same category. Location of facial marks are compared across images in barycentric coordi-

 1544

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

Fig. 11. (a) Potential facial marks detected by the multiscale automatic facial mark detector. The different colors indicate the number of scales at which the facial
marks are detected. Red is two levels, blue is three levels, green is four levels, and yellow is five levels. (b) Detections represented by circles represent potential
false positives and are removed using gradient-based SVD.

Fig. 12. Facial marks detected by the multiscale automatic facial mark detector for a pair of twins. The different colors indicate the number of scales at which the
facial marks are detected. Red is two levels, blue is three levels, green is four levels, and yellow is five levels.

nate system as described in Section VI. Similarity is computed
by formulating the matching process in terms of a weighted bipartite graph. The normalized similarity between two images
and is defined as

and
are weighting factors that correspond to the
and
number of levels at which the facial marks are detected.

(16)

We perform extensive experiments on both the manually
annotated facial marks and the automatically detected facial
marks. In addition to the experiments mentioned in Section VI
(Experiment 1, Experiment 2 and Experiment 3), we include

where the optimal matches are represented by a set
, and
,

VIII. EXPERIMENTAL SETUP AND RESULTS

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

1545

Fig. 13. Twins versus Twins and All versus All scenario used for comparing the target set and the query set.

Experiment 4 and Experiment 5, which evaluate the performance of multiscale automatic facial mark detector. These new
experiments are described as follows,
1) Experiment 4
a) Experiment Setup: Match the facial marks detected
by the multiscale automatic facial mark detector. The
query set and the target set are composed of detections from the automatic facial mark detector from
session-1 or session-2 dataset.
b) Target Set: Detections from Session- data, where
represents the session number and
or 2.
c) Query Set: Detections from Session- data
d) Inference: Performance is lower when compared to
Experiment 1 but it is more consistent with change in
scenarios.
2) Experiment 5
a) Experiment Setup: Match the facial marks detected
by the multiscale automatic facial mark detector. The
query set and the target set are composed of detections
from the automatic facial mark detector from different
sessions.
b) Target Set: Detections from Session-1 Data
c) Query Set: Detections from Session-2 Data
d) Inference: Consistency in performance is observed
when comparing facial marks across sessions, unlike
the trend observed with manual detections in Experiment 2 and Experiment 3.
For each of the aforementioned experiments, we define two
different scenarios (Twins versus Twins and All versus All) of
comparing the target set and the query set to generate performance curves. Fig. 13 depicts both of these scenarios. The main
difference between these two scenarios lies in the generation of
the impostor scores. For both scenarios the genuine scores are
obtained by comparing face images of a subject in the query set
against other images of the same subject in the target set. However, for the Twins versus Twins scenario, the impostor scores
are obtained by comparing a face image of a subject in the target
set only against the images of the subject’s twin. The impostor
scores for All versus All are obtained by comparing a face image
of a subject in the target set against images of all other subjects
in the query set.

Fig. 14. Distribution of moles, freckles, and acne across the first 50 subjects in
the dataset, i.e., 25 pairs of twins.

Finally, the last variation added only to Experiment 1 is that
it is executed for different subsets of facial mark categories provided. Each manually annotated facial mark is characterized
by its geometric location and a facial mark category. Amongst
the different categories of facial marks defined, moles, freckles
and pimple are more common and prominent. Therefore we define different subsets of facial mark categories to determine if
some facial mark categories are more useful compared to others.
Fig. 14 shows the distribution of moles, freckles and pimple for
the first 50 pairs of twins. The following subsets of facial marks
are also considered:
1)

2)
3)
4)

; i.e., each manually annotated facial mark is
characterized only by its geometric location and the facial
mark category is ignored.
Since the automatically detected facial marks are characterized only by the geometric location, we do not have to execute

 1546

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

Fig. 15. ROC curves of Twins versus Twins comparison for various sets of
facial marks annotated by observer 1 in session 1.

Fig. 16. Performance of All versus All comparison for various sets of facial
marks annotated by observer 1 in session 1.

the experiments based on the subset of facial marks. We use Receiver Operating Characteristic (ROC) curves and Equal Error
Rate (EER) as performance measures to evaluate the performance of the different experiments. If the ROC curve is closer
to the upper-left corner, it indicates a better performing system
and correspondingly, a lower value of EER.
A. Results: Experiment 1
Figs. 15 and 17 present the ROCs for Experiment 1 for session 1 and session 2 dataset for Twins vs Twins comparison
using different subsets of facial marks. Similarly, Figs. 16 and
18 present the ROCs for Experiment 1 for session 1 and session
2 dataset for All versus All comparison using different subsets
of facial marks. Considering subsets of facial marks does not
significantly improve performance. In fact, best performance
is achieved by considering all categories of facial marks. We
observe that there is no significant degradation in performance

Fig. 17. ROC curves of Twins versus Twins comparison for various sets of
facial marks annotated by observer 1 in session 2.

Fig. 18. Performance of All versus All comparison for various sets of facial
marks annotated by observer 1 in session 2.

when facial mark categories are ignored. This indicates that geometric location of facial marks is a robust feature that can be
used to differentiate between identical twins. A similar trend
is observed in both scenarios. An interesting result is observed
when comparing the performance between Twins versus Twins
comparisons and All versus All comparisons. Performance is
significantly better when comparing facial marks across unrelated individuals (All versus All scenario) with facial marks
across identical twins. This leads to an inference that distribution of facial marks across identical twins appears to be correlated.
The distributions of genuine and impostor scores for both experiments from session 1 are shown in Fig. 19. There exists a
larger overlap between the match and nonmatch scores for the
Twins versus Twins comparison compared to the All versus All
comparisons, leading to higher error and lower performance.
Hence, it is easier to differentiate unrelated persons using facial
mark distribution than identical twins. Although results are provided based on the facial marks annotated by observer 1, similar

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

1547

TABLE II
EQUAL ERROR RATES COMPUTED FOR TWINS VERSUS TWINS AND ALL VERSUS
ALL COMPARISONS ACROSS SESSIONS FOR OBSERVER 1 AND OBSERVER 2

Fig. 19. Distributions of match and nonmatch scores for Twins versus Twins
and All versus All comparisons.

Fig. 21. Performance curves for Experiment 3 for Twins versus Twins and All
versus All comparisons for observer 1 versus observer 3, session 1.

computed for both twin and unrelated persons comparison
across sessions for observer 1 and observer 2.
C. Results: Experiment 3

Fig. 20. Comparing the performance curves of observer 1 for both All versus
All and Twins versus Twins comparisons, for Experiment 2.

trends are seen in the case of facial marks annotated by other observers.
B. Results: Experiment 2
In Experiment 2 we compare manually annotated facial
marks across sessions. Only two observers, observer 1 and
observer 2, took part in both the sessions. Hence, we can
compare annotations made only by them. Fig. 20 presents the
ROC curves for both the Twins versus Twins and All versus
All comparison considering all facial marks (FM). Again, we
observe that the performance of All versus All outperforms
the Twins versus Twins performance. Ideally, the performance
should be similar to the results of Experiment 1. However,
there is a degradation in overall performance when comparing
manually detected facial marks across sessions, indicating that
observers are not consistent over time. Table II lists the EER

In this experiment, similarity scores are computed by comparing a query set annotated by one observer to a target set annotated by a different observer. Performance curves for these
experiments for each session are shown in Fig. 21 for observer
1 versus observer 3 and Fig. 22 for observer 1 versus observer
4, for both Twins versus Twins and All versus All scenarios.
The performance degrades considerably when facial marks annotated by different observers are compared against each other.
This occurs due to the variation of facial mark annotations by
observers. There is lack of uniformity in facial mark annotations
across observers. Though all observers annotated the prominently visible facial marks, a few failed to annotate the less
prominent marks. However, even in these experiments, the performance of All versus All comparisons is better than the Twins
versus Twins suggesting similarity in the distribution of facial
marks across twins compared to unrelated persons.
D. Results: Experiment 4 and Experiment 5
Experiment 4 and Experiment 5 evaluate the performance of
the proposed multiscale automatic facial mark detector. A total
of 4602 facial marks were detected in the session 1 dataset and
3012 facial marks were detected in the session 2 dataset. Figs. 23
and 24 shows the performance curves for Experiment 4 for both
the Twins versus Twins and All versus All comparisons for each

 1548

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

Fig. 22. Performance curves for Experiment 3 for Twins versus Twins and All
versus All comparisons for observer 1 versus observer 4, session 2.

Fig. 24. ROC curves representing the performance of the proposed multiscale
automatic facial mark detector for Experiment 4, for session 2 dataset, for both
Twins versus Twins and All versus All comparison.

Fig. 23. ROC curves representing the performance of the proposed multiscale
automatic facial mark detector for Experiment 4, for session 1 dataset, for both
Twins versus Twins and All versus All comparison.

Fig. 25. ROC curves representing the performance of the proposed multiscale
automatic facial mark detector for Experiment 5, for both Twin versus Twin and
All versus All comparison.

of the dataset. The performance of the proposed detector is relatively worse when compared with the results obtained for Experiment 1, which uses manual annotations. The performance
curves for Experiment 5 are shown in Fig. 25. Facial marks detected in session 2 dataset are compared against facial marks detected in session 1 dataset. In theory, this is analogous to Experiment 2. We obtain similar results for both Twins versus Twins
comparisons and All versus All comparisons in Experiment 4
and Experiment 5. It does not exhibit a large variation in performance as seen in Experiment 2 and Experiment 3.
Finally, we compare the performance between manually
detected facial marks and automatically detected facial marks.
Fig. 26 compares performance across Experiment 1, Experiment 2, Experiment 4, and Experiment 5 for the Twins versus
Twins comparison. Although the best performance is achieved
by comparing facial marks annotated by an observer at a given
time, there is a large degradation in performance when we

compare facial marks annotated by an observer over time,
indicating inconsistency in performance. This variation in
performance is not seen when comparing facial marks detected
by the proposed multiscale automatic facial mark detector.
The automatic facial mark detector performs better than facial
marks annotated by an observer at different instances of time.
Similarly, we observe that the performance of the automatic
facial mark detector is better when compared with Experiment
3, i.e., when we compare facial marks annotated by different observers, as shown in Fig. 27. The performance of the proposed
automatic facial mark detector is relatively consistent and uniform. The same trend is observed in All versus All comparison.
The results obtained can be summarized as follows,
1) In every experiment, we observe that the All versus All
comparison performs better than Twins versus Twins comparison. This indicates that the facial mark distributions
across identical twins appear to be correlated.

 SRINIVAS et al.: ANALYSIS OF FACIAL MARKS TO DISTINGUISH BETWEEN IDENTICAL TWINS

1549

experiments are designed and implemented to highlight the performance of the proposed automatic facial mark detector when
compared with the performance achieved by using manually detected facial marks. From the results there appears to be a correlation in distribution of facial marks across twins. This phenomenon is observed across all experiments. In the future, we
will explore using richer facial mark characteristics like texture,
shape and color to improve performance. We hope to further explore the use of different matching algorithms and compare it
to the proposed matching algorithm. Facial marks features can
be fused with other facial features to enrich facial characterizations for improved performance. The results of the investigation
makes a case for the use of facial marks in biometric characterization.
REFERENCES
Fig. 26. Performance comparison between manually detected facial marks experiments (Experiment 1 and Experiment 2), and automatically detect facial
mark experiment (Experiment 4 and Experiment 5). For Twins versus Twins
comparisons.

Fig. 27. Performance comparison between manually detected facial marks experiment (Experiment 3), and automatically detect facial mark experiment (Experiment 4). For Twins versus Twins comparisons.

2) Manual annotation process is a difficult and time consuming task, and hence there is a need for a robust
automatic facial mark detector.
3) Though the performance of the proposed multiscale automatic facial mark detector is slightly lower than the performance of manual annotations for a single session, it exhibits greater consistency over time.
IX. CONCLUSION
In this paper, we analyzed the usefulness of facial marks as
a potential biometric signature for distinguishing between identical twins. We proposed a multiscale automatic facial mark detection system for distinguishing between identical twins solely
based on the geometric distribution of facial marks. Different

[1] Z. Sun, A. A. Paulino, J. Feng, Z. Chai, T. Tan, and A. K. Jain, “A study
of multibiometric traits of identical twins,” in Proc. SPIE, Biometric
Technology for Human Identification VII, Apr. 2010, pp. 1–12.
[2] K. W. Bowyer, “What surprises do identical twins have for identity
science?,” Computer, vol. 44, pp. 100–102, Jul. 2011.
[3] P. Phillips, P. Flynn, K. Bowyer, R. Voder Bruegge, P. Grother, G.
Quinn, and M. Pruitt, “Distinguishing identical twins by face recognition,” in Proc. IEEE Int. Conf. Automatic Face Gesture Recognition
and Workshops, 2011, Mar. 2011, pp. 185–192.
[4] A. K. Jain, S. Prabhakar, and S. Pankanti, “On the similarity of identical
twin fingerprints,” Pattern Recognit., vol. 35, no. 11, pp. 2653–2663,
2002.
[5] J. Daugman and C. Downing, “Epigenetic randomness, complexity and
singularity of human iris patterns,” in Proc. Biological Sciences, 2001,
vol. 268, no. 1477, pp. 1737–1740.
[6] W. Zhao, R. Chellappa, P. J. Phillips, and A. Rosenfeld, “Face recognition: A literature survey,” ACM Computer Survey, vol. 35, pp. 399–458,
Dec. 2003.
[7] D. Lin and X. Tang, “Recognize high resolution faces: From macrocosm to microcosm,” in Proc. IEEE Computer Society Conf. Computer
Vision and Pattern Recognition, 2006 (CVPR’06), 2006, vol. 2, pp.
1355–1362.
[8] [Online]. Available: http://skin-care.health-cares.net/skin-lesions.phpLast accessed: 07/2010
[9] G. Loy and A. Zelinsky, “A fast radial symmetry transform for
detecting points of interest,” in Proc. 7th Eur. Conf. Computer
Vision—Part I (ECCV’02), London, U.K., 2002, pp. 358–368,
Springer-Verlag.
[10] T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham, “Active
shape models-their training and application,” Comput. Vis. Image
Understand., vol. 61, pp. 38–59, Jan. 1995.
[11] C. Bradley, The Algebra of Geometry: Cartesian, Areal and Projective
Co-Ordinates. Bath: Highperception, 2007.
[12] G. Zhu, D. L. Duffy, A. Eldridge, M. Grace, C. Mayne, L. O’Gorman,
J. F. Aitken, M. C. Neale, N. K. Hayward, A. C. Green, and N. G.
Martin, “A major quantitative-trait locus for mole density is linked
to the familial melanoma gene cdkn2a: A maximum-likelihood combined linkage and association analysis in twins and their sibs,” Amer.
J. Human Genetics, vol. 65, pp. 483–492, Aug. 1999.
[13] N. Srinivas, G. Aggarwal, P. Flynn, and R. Voder Bruegge, “Facial
marks as biometric signatures to distinguish between identical twins,”
in Proc. IEEE Computer Society Conf. Computer Vision and Pattern
Recognition Workshops (CVPRW 2011), Jun. 2011, pp. 106–113.
[14] J. S. Pierrard and T. Vetter, “Skin detail analysis for face recognition,”
in Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2007
(CVPR’07), Jun. 2009, pp. 1–8.
[15] Z. Zhang, S. Tulyakov, and V. Govindaraju, “Combining facial skin
mark and eigenfaces for face recognition,” in Proc. Third Int. Conf.
Biometrics (ICB’09), Berlin, Heidelberg, Germany, 2009, pp. 424–433,
Springer-Verlag.
[16] A. Jain and U. Park, “Facial marks: Soft biometric for face recognition,” in Proc. 16th IEEE Int. Conf. Image Processing (ICIP 2099),
Nov. 2009, pp. 37–40.
[17] A. W. K. Kong, D. Zhang, and G. Lu, “A study of identical twins palmprints for personal verification,” Pattern Recognit., vol. 39, no. 11, pp.
2149–2156, 2006.

 1550

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012

[18] E. W. K. Kodate, R. Inaba, and T. Kamiya, “Facial recognition by a
compact parallel optical correlator,” Meas. Sci. Technol., vol. 13, no.
11, p. 1756, Nov. 2002.
[19] American National Standard for Information Systems-Data Format
for the Interchange of Fingerprint, Facial, and Other Biometric Information, Part 2: Xml Version [Online]. Available: http://www.nist.gov/
itl/ansi/upload/Approved-XML-Std-20080828.pdf, Last accessed:
07/2010
[20] S. Milborrow and F. Nicolls, “Locating facial features with an extended
active shape model,” in Proc. Eur. Conf. Computer Vision (ECCV),
2008, pp. 504–513 [Online]. Available: http://www.milbo.users.sonic.
net/stasm
[21] H. W. Kuhn, “50 years of integer programming 1958–2008,” in The
Hungarian Method for the Assignment Problem. Berlin, Heidelberg,
Germany: Springer, 2010, pp. 29–47.
[22] P. Burt and E. Adelson, “The Laplacian pyramid as a compact image
code,” IEEE Trans. Commun., vol. 31, no. 4, pp. 532–540, Apr. 1983.
[23] D. Reisfeld, H. Wolfson, and Y. Yeshurun, “Context free attentional
operators: The generalized symmetry transform,” Int. J. Comput. Vis.,
vol. 14, pp. 119–130, 1995.
[24] K. Carolyn, D. Ballard, and J. Sklansky, “Finding circles by an array
of accumulators,” Commun. ACM, vol. 18, pp. 120–122, Feb. 1975.
[25] X. Zhu and P. Milanfar, “A no-reference sharpness metric sensitive
to blur and noise,” in Proc. Int. Workshop on Quality of Multimedia
Experience, 2009., Jul. 2009, pp. 64–69.
[26] X. Feng and P. Milanfar, “Multiscale principal components analysis
for image local orientation estimation,” in Proc. Thirty-Sixth Asilomar
Conf. Signals, Systems and Computers, 2002., Nov. 2002, vol. 1, pp.
478–482.
Nisha Srinivas is a Graduate Research Assistant enrolled in the Ph.D. program in computer science and
engineering at the University of Notre Dame. She received the M.S. degree in electrical engineering from
Syracuse University in 2009 and the B.E. degree in
electronics and communication from Bangalore Institute of Technology, Bangalore, in 2006. Her current work focuses on studying the usefulness of facial
marks as biometric signatures to distinguish between
individuals.

Gaurav Aggarwal is a scientist with Yahoo! Labs
Bangalore. He received the B.Tech. degree in computer science and engineering from the Indian Institute of Technology, Madras, in 2002, and the M.S.
and Ph.D. degrees in computer science from the University of Maryland, College Park, in 2004 and 2008,
respectively.
He has held a Research Scientist position with Object Video Inc. and a Research Assistant Professor
position with the Department of Computer Science
and Engineering, University of Notre Dame before
joining Yahoo. His research interests are in image and video processing, computer vision, pattern recognition, and machine learning.

Patrick J. Flynn (F’12) is Professor of Computer
Science and Engineering and Concurrent Professor
of Electrical Engineering at the University of Notre
Dame. He received the B.S. degree in electrical
engineering (1985), the M.S. degree in computer
science (1986), and the Ph.D. degree in computer
science (1990) from Michigan State University, East
Lansing.
He has held faculty positions at Notre Dame
(1990–1991, 2001–present), Washington State
University (1991–1998), and Ohio State University
(1998–2001). His research interests include computer vision, biometrics, and
image processing.
Dr. Flynn is an IAPR Fellow, an ACM Distinguished Scientist, a past
Associate Editor-in-Chief of IEEE TRANSACTIONS ON PATTERN ANALYSIS AND
MACHINE INTELLIGENCE, and a past Associate Editor of IEEE TRANSACTIONS
ON INFORMATION FORENSICS AND SECURITY, IEEE TRANSACTIONS ON IMAGE
PROCESSING, IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE
INTELLIGENCE, Pattern Recognition, and Pattern Recognition Letters. He has
received outstanding teaching awards from Washington State University and
the University of Notre Dame.

Richard W. Vorder Bruegge is the Senior Photographic Technologist at the Federal Bureau of
Investigation (FBI), in the Digital Evidence Laboratory, where he is responsible for overseeing science
and technology developments in the imaging sciences. He received the B.Sc. degree in engineering
(1985), the M.Sc. degree in geological sciences
(1987), and the Ph.D. degree in geological sciences
(1991) from Brown University.
He has been with the FBI since 1995, where he has
performed forensic analysis of image and video evidence, testifying in state, federal, and international courts as an expert witness
over 60 times. His research interests include the forensic analysis of image evidence, with a particular interest in face and iris recognition.
Dr. Vorder Bruegge was chair of the Scientific Working Group on Imaging
Technology (SWGIT) from 2000 to 2006 and is current chair of the Facial Identification Scientific Working Group (FISWG). He is a fellow of the American
Academy of Forensic Sciences (AAFS) and in 2010 he was named a Director
of National Intelligence (DNI) Science and Technology Fellow.