(2004) Effects of Illumination Changes on the Performance of Geometrix FaceVision 3D FRS

Effects of Illumination Changes on the Performance of
Geometrix Facevision@3D FRS
Eric P. Kukula Stephen 1. Elliott, PhD Roman Waupotitsch Bastien Pesenti
kukula@ purdue.edu ellion@purdue.edu romanw@geometrix.com bastienp@geometrix.com
Industrial Technology, Industrial Technology, Vice President of R&D Research Engineer
Purdue University Purdue University Geometrix Inc Geometrix Inc
West Lafayette, IN 47906, West Lafayene, IN 47906, 1590 The Alameda Ste 200 1590 The Alameda Ste 200
USA USA San Jose. CA 95 124 San Jose, CA 95 124

ABSTRACT environmental conditions, such as lighting, may be
inconsistent, consequently affecting the
This evaluation examined the effects of four performance of the face recognition system. In
frontal light intensities on the performance of a 3D previous research by Kukula and Elliott [l], a
face recognition algorithm, specifically testing the commercially off the shelf software (COTS) 2D
significance between an unchanging enrollment facial recognition algorithm was assessed, which
illumination condition (220-225 lux) and four revealed that 2D face recognition still has
different illumination levels for verification. The significant challenges to overcome with regard to
evaluation also analyzed the significance of external illumination, specifically when the ambient lighting
artifacts (i.e. glasses) and personal characteristics is low, as well as when light was not held constant.
(i.e. facial hair) on the performance of the face Recently, three dimensional face recognition
recognition system (FRS). algorithms have started to emerge in the
Collected variables from the volunteer crew marketplace. According to the manufacturers, 3D
included age, gender, ethnicity, facial face recognition has advantages over 2D face since
characteristics, hair covering the forehead, scars on it compares the 3D shape of the face, which is
the face, and glasses. invariant in different lighting conditions and pose,
The analysis of data revealed that there are no although light conditions were only evaluated in this
statistically significant differences between evaluation.
environmental lighting and 3D FRS performance Over the past ten years there have been three
when a uniform or constant enrollment illumination large scale independent evaluations conducted on
level is used. 2D COTS facial recognition systems which have
shown that performance dramatically decreases
Keywords: biometrics, 3D face recognition, when environment lighting changes [ 2 4
environmental conditions, performance testing Currently, independent testing of 3D systems is
sparse as it is an emerging biometric technology.
MOTIVATION However, intemal testing conducted by Geometrix
have reported equal error rates (EER) of less than
As govemment and private corporations begin 2% using image databases from University of
to implement biometric technologies in operational Southern California and the University of Notre
settings, such as in airports and facility access Dame. At the time of writing, no independent
control, the environment and application must be testing of COTS 3D face recognition has been
fully examined before implementation. With regard complete. However the NIST Face Recognition
to face recognition, there are several challenges to Grand Challenge (FRGC) is currently underway
face recognition systems, including illumination, with report set to be released in August of 2005.
which may affect the performance of the system. Further internal studies of the Geometrix Face
The implementation of biometric systems, including Vision system commissioned by the Defense
face recognition systems into legacy environments Advanced Research Projects Agency (DARPA)
that may not have ideal environmental conditions, concluded that as little as 6 gray values are
4
indicate that this is an area of research that is sufficient for the Facevision system to perform
important as deployments of face recognition high-quality 3D reconstruction of faces. However,
systems become pervasive. As a result until now no independent performance assessments

331 02004 IEEE
0-7803-8506-3/02/$17.00

Authorized licensed use limited to: Purdue University. Downloaded on February 27,2010 at 12:03:45 EST from IEEE Xplore. Restrictions apply.

using different lighting conditions have been
S l o i - Zd’aflIhegrauird
performed. The purpose of the evaluation reported 3 Light sources . a s ’ /mm ground Io the Mitom 01 the enclosure
here was designed to address exactly this aspect of
3D recognition, namely to perform a system-level
test of the Geometnx Facevision system.

CONCEPT OF THE SYSTEM

The 3D face recognition system used in this
evaluation was the Geometrix Human Identification
System (HIS). The system’s fundamental algorithms
were inspired by Chen and Medioni [ 5 ] .The sensor
used was the Face Vision 200, which captures two
images using two stereo calibrated cameras. The
system then processes the images using proprietary
and patented algorithms to construct a metrically
accurate 3D model of the face.
The 3D face model is then further processed to Figure 1: Testing Environment
create a fully textured version of the face that may
be used for visual inspection by an operator.
Moreover, a 3D face template is extracted from the A light impermeable curtain segregated the testing
model [6,7], which is 3 kilobytes for one-to-one environment from the educational computer lab. All
verification and less than 200 bytes for one-to-many fluorescent lighting was removed from the testing
identification. Verification time in this evaluation environment and the curtain impeded the
averaged 12 seconds on a single processor, while uncontrolled fluorescent illumination from the
intemal testing using dual processors averaged less educational lab area, resulting in a stable zero
than 6 seconds. illuminance (lux) environment. The background
The 3D face template encodes the salient used was very close to the recommended 18% gray
features of the face with patented Active FusionTM [9-IO]. The extemal lighting used for verification
algorithms, which allows a very accurate composed of three JTL Everlight continuous
comparison between the “enrollment” face with the halogen lamps with 500 Watt USHIO halogen bulbs
captured “verification” face. Robustness techniques covered by 24 inch softboxes. The lamps were
are used to weigh different aspects of the face positioned in a manner that created an evenly
according to their contribution to “being able to illuminated face. The Geometrix Facevision 200
distinguish two faces” and their robustness to camera system, shown in included a lighting system
changes in the facial shape over time and changes that remained constant throughout the evaluation
due to facial expression, which were both outside (both enrollment and verification). The illumination
the scope of this study. of the experimental area was monitored with a NIST
certified broad range ludfc light meter. The Face
SETUP
EXPERIMENTAL Vision 200 camera system included two off-the-
shelf USB cameras. The cameras were attached to a
This evaluation took place in the Biometric Dell Omniplex GX260 computer through an Orange
Standards, Performance, and Assurance Laboratory Micro USB 2.0 PCI card. The computer was a
in the School of Technology at Purdue University. single 2.0 GHz processor, 512 MB RAM, 40 GB
The testing environment, shown in Figure 1, was hard drive. The operating system was Microsoft
similar to that of Blackbum, Bone, and Phillips [8] Windows XP Pro SPl.
and the setup described by Kukula and Elliott [1,8].

332


The COTS unit is currently optimized to capture
faces between 18 inches and 30 inches for
enrollment, and 16 inches to 36 inches for
verification or identification.
The sensor was originally calibrated by
Geometrix. On-site color and sensitivity calibration
was performed once in the Biometrics Standards,
Performance, and Assurance Laboratory to optimize
the sensor in the environment. It was subsequently
Figure 2: Geometrix Facevision 200 camera system inspected each day in accordance with the testing
Lighting protocol.

This evaluation tested the performance of a 3D Software
face recognition algorithm using one enrollment
lighting intensity and four verification lighting Geometrix provided all software that was used
intensities. The enrollment lighting intensity used in this evaluation. The 3D model creator was
only the Geometrix system LED lights, which were Facevision 200 Series v5.1. The evaluation also
fastened to each side of the camera mount, which used the Geometrix Facevision Human
can be seen in Figure 2. The illumination defined for Identification System (Facevision HIS) version
enrollment was 220 - 225 lux. These LED lights 2.3. The system provides both an interface for
remained on throughout testing. Verification enrollment and verification or identification
occurred at 4 different light intensities as described
operations, as well as administrative tools to
in Table 1.
manage the database of enrolled persons
Table 1: Definition of lighting conditions (Figure 3). However for this evaluation only the
Use I Name I Light Intensity enrollment and verification software was used.
I Light Condition 1 I 220-225 lux
U
Enrollment/
Verification
FACEVISION
Verification Light Condition 2 320-325 lux HIS GUI
Verification Light Condition 3 650-655 lux
Verification Light Condition 4 1020-1140 lux

Hardware

The COTS Geometrix Facevision FV200
sensor was used (Figure 2) for image acquisition. It
is a passive stereo-based sensor incorporating hoard-
level cameras and custom lenses, which is
connected to a computer using a USB 2.0 interface.
The dimensions of the sensor are approximately
FACEVISION
6.5x4.3x2.5 inches. This sensor was used for both
FVZOO SENSOR saL SERVER
enrollment and verification. The Facevision 200 Copyright 0 GEOMETRIX
sensor incorporates an LED based lighting unit that
is attached on each side of the system. The lights are Figure 3: Facevision HIS
dimmable. However, when set at the recommended The enrollment mode is designed to enroll new
intensity (220-225 lux), the LED light system persons, add additional biometric templates for
provides sufficient illumination for the sensor to existing persons, and access or edit demographic
operate in an optimal manner, even in the darkest information. The Facevision HIS software provides
environment. For the purpose of this evaluation, the a seamless interface for operating the Facevision
protocol called for the system lights to remain at the FV200 capture sensor. While the enrollment process
recommended level of 220-225 lux throughout the is fully automatic, a manual step may be performed
experiment. to verify the enrollment data. This step was

333


performed during each enrollment as part of of the image labeled light condition 2 was 1.631.625
protocol, in order to verify model quality. or 2.608.
The verification mode is designed to verify the
claimed identity of the captured person with the 3D
template stored in the database. ABer a few seconds,
the system gives a binary answer, “Access Granted
or “Access Denied.” The system also displays a
confidence rating of the decision made, as well as a
list of potential impostors known in the system.
However, only the binary response was used for
data analysis in this evaluation.

Captured Image Specifications

To eliminate extemal effects on the experiment
and to emphasize the sole effect of lighting on the
performance of the system, the subject’s position,
facial pose, and face covering artifacts were defined
by the test protocol. Specifically, faces were
captured with the nose approximately centered in
the image. To simplify the process, each participant
remained seated during the evaluation two feet from
Light Condition 3
the ground. To compensate for the varying heights
of participants, the camera was attached to a
mechanical tripod that could be adjusted in height. Figure 5: Sample images from the 4 tested light
The resulting captured image reflects the proposed intensities
face recognition data format specification for
captured images [7], which can be seen in Figure 4.
EVALUATION
CLASSIFICATION
This document suggests the image should be
The evaluation was defined as cooperative,
overt, unhabituated, attended, and closed [ I 11. The
experimental evaluation is classified as a modified
technology evaluation. A traditional technology
evaluation is conducted in a laboratory, by a
universal sensor, and using the same data causing
repeatability of samples. In this case however, data
was collected and was evaluated on-line with the
specific results and scores presented after the
completion of the computation, hence its
classification a s a modified technology evaluation.
The purpose of the evaluation was to assess the
t!A
effects of four frontal light intensities. Failure to
Figure 4: WCITS face recognition data format Enroll, Failure to Acquire, and a statistical analysis
image requirement (Griffin, 2003) of the differences in light and performance of the
device were assessed.
centered, meaning the mouth and middle of the nose
should lie on the imaginary line AA (Figure 4). The Volunteer Crew
location of the eyes in the images should range
between 50-70% the distance from the bottom of the This evaluation involved thirty subjects from
image and the width of head ratio ( N C C ) should be the School of Technology at Purdue University.
no less than 7/4 (1.75). Images collected in this Demographic information can be seen in Table 2.
study fully conformed to the requirements proposed
in [9], as seen in Figure 5. The width-to-head ratio

334


Table 2: Volunteer crew demographic information distance used for both enrollment and verification in
this evaluation was 28 inches. To monitor the light,
subjects were asked to hold a light meter sensor in
front of their nose periodically throughout the
evaluation to monitor the lighting conditions. These
readings were recorded and checked to maintain
Caucasian 24
repeatability throughout the study.
African
The generalized testing protocol model can be
American 1
seen in Figure 6. This evaluation was designed to
Asian 2
compare the stored 3D face template created in the
Hispanic 3
enrollment lighting condition (220-225 lux) against
Native
verification attempts captured at the four different

I 30-39
40-49 3

no 26

Yes 7
no 23

Figure 6: Protocol Design
TESTING
PROTOCOL
light intensities: 1) Enrollment lighting (220-225
lux), 2) Light condition 2 (320-325 lux), 3) Light
The protocol used for this evaluation called for condition 3 (650-655 lux), and 4) Light condition 4
calibration of the cameras each day testing occurred. (1020-1 140 lux). The protocol called for 3
At this time the operator also verified the verification attempts in each of the four light
experimental setup of all the equipment used for the
intensities, for a total of 12 attempts for each
study. The testing protocol consisted of one subject.
enrollment light condition and four Verification light
conditions. The lighting conditions are defined in
Enrollment
Table 1. Before data collection began, participants
were informed of the testing procedures and given The first testing procedure was enrollment.
specific instructions, which included:
After the subject was seated and the camera position
Remove eyeglasses, hats, or caps
was verified, the test operator notified the subject
Refrain from chewing gum or candy the image capture sequence was beginning. During
Look directly at the sensor (between the two this sequence music could be heard. After the
cameras) and maintain a neutral expression capture sequence was complete, a 2D image
Stay as still as possible while the music is appeared which was checked for quality (no facial
playing. expressions, closed eyes, etc). The three
At this time, the field of view of the camera was dimensional model was then computed, checked for
checked to ensure captured images resembled correct nose position and quality, then stored. An
Figure 4. The distance between the camera and the example of a 3D model used in this study is shown
test subject's face was also measured to ensure the in Figure 7.
proper camera depth of field was achieved. The

335


mode when presented with the various light levels.
A statistical analysis shows that at an alpha level of
0.01, there was no statistically significant difference
in the performance of the algorithm when the light
level was measured between light level 1 (220-225
lux), and the other levels (320-325 lux; 650-655 lux;
1020-1140 lux).

CONCLUSION

Because the Geometrix face recognition engine uses
Figure 7: Example of a 3D model a template extracted from 3D, unlike the 2D image-
Verification based engines, this study shows that this 3-D
algorithm seems to have overcome the usual
Verification followed the same procedure for limitations of illumination variations. Unlike a
each subject. The light conditions followed a previous study [I], this evaluation has shown that
structured order and were not randomized. After there are no statistically significant differences in
enrollment was complete, 3 verification attempts performance at any of the tested illumination levels.
were conducted in the same lighting intensity used Further research is underway to evaluate lighting
for enrollment (light condition l), followed by 3 angles and pose, to establish the progress of 3-D
attempts in light conditions 2, 3, and 4. Figure 8 face recognition algorithms.
shows the visual display given to the operator after
each verification attempt. To ensure data collection REFERENCES
was accurate a screen shot of each attempt was
collected, that used a barcode naming convention Kukula, E., & Elliott, S. (2003). Securing a
that removed data collection errors of keying data, Restricted Site: Biometric Authentication at
as well as reduced time between attempts. Entry Point. Paper presented at the 37th
Annual 2003 Intemational Camahan
Conference on Security Technology (ICCST),
(pp. 435439), Taipei ,Taiwan, ROC

Phillips, P., Rauss, P., & Der, S. (1996).
FERET (Face Recognition Technology)
Recognition Algorithm Development and Test
Report (ARL-TR-995): U.S. Army Research
Laboratory.

Blackbum, D., Bone, J., & Phillips P. (2000).
Face Recognition Vendor Test (FRVT) 2000
y_
Evaluation Report. DoD, DAFVA, NIJ
Figure 8: Feedback from a verification attempt Phillips, P., Grother, P., Bone, M., Micheals,
R., Blackbum, D., & Tabassi, E. (2003). Face
RESULTS Recognition Vendor Test 2002: DARF'A,
NIST, DOD, NAVSEA.
The study consisted of 30 individuals, 30
enrollment attempts, 30 impostor attempts and 360 Chen G. & Medioni G., Building Human Face
genuine verification attempts. At the enrollment Models from Two Images, Journal of VLSI
stage there were no failure to enrolls (FTE = O%), or Signal Processing, Kluwer Academic
Failure to Acquires (FTA = 0%). Publishers, vol. 27, no. U2, pp. 127-140,
The hypotheses were set up to establish whether January 2001
there was any significant difference in the
performance of the algorithm in the verification

336


Waupotitsch R. & Medioni G. Robust
Automated Face Modeling and Recognition
Based on 3 0 Shape. Biometrics Symposium
on Research, Crystal City, September 2003.

Waupotitsch R. & Medioni G. Face Modeling
and Recognition in 3 - 0 . AMFG 2003: 232-
233.

Bone, M. and D. Blackbum, Face Recognition
at a Chokepoint: Scenario Evaluation Results.
2002, DoD Counterdrug Technology
Development Program Office: Dahlgren. p.
58.

Griffin, P. (2003). Face Recogntion Format
for Data Interchange (M1/04-0041): INCITS
M1.

Rubenfeld, M., & Wilson, C. (1999). Gray
Calibration of Digital Cameras To Meet NIST
Mugshot Best Practice. NIST IR-6322

[ l l ] Mansfield, A.J. and J.L. Wayman, Best
Practices in Testing and Reporling
Performances o Biometric Devices. 2002,
f
Biometric Working Group. p. 32.

337


(2004) Effects of Illumination Changes on the Performance of Geometrix FaceVision 3D FRS

More Related Content

What's hot (19)

Similar to (2004) Effects of Illumination Changes on the Performance of Geometrix FaceVision 3D FRS (20)

More from International Center for Biometric Research (20)

Recently uploaded (20)

(2004) Effects of Illumination Changes on the Performance of Geometrix FaceVision 3D FRS