Transcript
ICIP 2016 COMPETITION ON MOBILE OCULAR BIOMETRIC RECOGNITION A. Rattani1, R. Derakhshani1, S. K. Saripalle2 and V. Gottemukkula2 1University of Missouri- Kansas City, USA 2EyeVerify Inc., USA
Outline • Introduction to Ocular Biometrics • Existing Databases on Ocular Biometrics • VISOB ICIP 2016 Challenge Version • Database Description • Protocol • Participating Teams • Results by Organizers • Conclusions
Introduction to Ocular Biometrics • Ocular biometrics: using characteristic features extracted from the eyes and immediate vicinity for personal recognition.
• Ocular biometric modalities have mainly focused on the iris, conjunctival vasculature, and periocular regions.
• Iris recognition is mostly performed in near infrared spectrum.
• Conjunctival vasculature and periocular regions are mostly imaged in the visible spectrum.
Figure 1: Example Eye Image with Iridial, Conjunctival, and Periocular regions labeled
Mobile Ocular Biometrics • With the unprecedented popularity of mobile devices and services, biometric technologies are being integrated into mobile devices at a fast pace
• The most natural user interactions with their mobile devices are either by touching them or looking at them
• Fingerprints and ocular biometrics are thus arguably the most important mobile biometric modalities
Contd.., • Ocular recognition in the visible spectrum is also an important research area because: No-touch recognition at a distance
Implementation with existing RGB cameras, especially in case of conjunctival and periocular-based modalities Inherent high information density and certain aspects of privacy
Contd.., • Ocular biometrics in the visible spectrum is rather nascent. • Unavailability of standards and publicly available large scale mobile ocular biometric databases • There are few publicly available databases on ocular biometrics such as UPOL, UBIRIS and MICHE, however: • They do not facilitate large-scale evaluation of the recognition schemes, • They do not necessarily provide eye images captured by mobile devices
Contributions 1. We have collected a large scale VISible light Ocular Biometric (VISOB) database of about 550 subjects using mobile phone cameras.
• VISOB comprises of eye images captured using front facing (selfie) cameras of three different mobile devices, namely Oppo N1 Samsung Note 4 iPhone 5s
Contd.., • VISOB dataset presents possible intra-class variations such as out-of-focus images occlusions due to prescription glasses different lighting conditions natural gaze deviations eye-makeup (i.e., eye liner and mascara) specular reflections and motion blur Other imaging artifacts
Contd.., 2. We conducted an international competition on VISOB through ICIP 2016.
• Overall aims: • large scale evaluation of biometric recognition schemes using visible light “selfie” captures of ocular regions • promote further research and advancement in this important area
Existing Databases on Visible Ocular Recognition
Figure 1: UPOL database Size: (384 x 64 volunteers) Acquisition system: TOPCON TRC50IA Remarks: Limited to high quality iris images, segmentation is facilitated by an encircling mask
Figure 2: UBIRIS v1 and v2 database Size: (1877 x 241 v1) Size: (11102 x 261 v2) Acquisition system: Nikon E5700 dSLR (v1) and Canon EOS 5D dSLR (v2) Competition: NICE I (robust iris segmentation) and NICE II (recognition schemes) competitions
Figure 3: UBIPr Size: (10252 x 344) Acquisition system: Canon EOS 5D dSLR Remarks: Mostly acquired for evaluating periocular recognition
Figure 3: MICHE I and II Size: 1600 x 50 MICHE I Size: 3120 x 75 MICHE II Acquisition system: iPhone 5s and Samsung Galaxy S4 (front and rear camera) Competition: MICHE I and MICHE II competitions (Mobile Iris Challenge Evaluation)
Figure 4: VSSIRIS database Size: (56 x 28 volunteers) Acquisition system: iPhone 5s and Nokia Lumia 1020 (rear camera)
VISOB Dataset: ICIP 2016 Challenge Version • Size: ~550 volunteers with an average of 10 images per eye (550 in original VISOB). • Acquisition device: iPhone 5s, Samsung Note 4 and Oppo N1. • Resolution: • iPhone set to capture bursts of still images at 720p resolution. • Samsung and Oppo devices capture bursts of still images at 1080p resolution. • Oppo has a higher end, AF selfie camera
• Volunteer’s data were collected during two visits (VISIT 1 and VISIT 2), 2 to 4 weeks apart.
Contd.., • Collection procedure: • At each visit, volunteers were asked to take selfie like captures using front facing cameras in two sessions. • Session 1 and Session 2 were about 10 to 15 minutes apart. • The volunteers used the mobile phones naturally, holding the devices 8 to 12 inches from their faces.
Contd.., • For each session, there are three lighting conditions: regular office light, dim office light, and natural daylight settings.
• Post-processing: • The collected database was cropped to retain only the eye regions of size 240 x 160 pixels using a Viola-Jones based eye detector.
VISOB Dataset Sample Images
Figure 5: Sample iPhone images captured in regular office light.
Figure 6: Sample iPhone images captured in daylight.
Figure 7: Sample iPhone images captured in dim office light.
Experimental Protocol • For Participants: VISIT 1 (550 subjects with 10 samples per subject) • Enrollment : Session 1 • Verification: Session 2 • Available only to the Organizers: VISIT 2 (290 subjects, from VISIT 1, with 10 samples per subject)
• Enrollment: Session 1 • Verification: Session 2
Mobile Device
Enrollment Set (# of images)
Verification Set (# of images)
VISIT 1 (# of images) iPhone
14077
13208
Oppo
21976
21349
Samsung
12197
12240
VISIT 2 (# of images) iPhone
12222
11740
Oppo
10438
9857
Samsung
9284
9584
Table 1. Enrollment and Validation stats for VISIT 1 and VISIT 2 of VISOB Dataset, ICIP2016 Challenge Version.
Participating Teams • Four universities and an industry participant submitted five algorithms to this competition. • The participants include: 1.
Norwegian Biometrics Laboratory, Norwegian University of Science and Technology (NTNU), Norway.
2.
Australian National University (ANU), Australia.
3.
Indian Institute of Information Technology Guwahati (IIITG), India and IBM Research India
4.
Anonymous (anonymized per participant’s request).
Participant
Features
Learning-based classifier
Non-learning based classifier
NTNU- 1
Maximum Response Filters
Deep Neural Networks
NA
NTNU- 2
Deep Sparse Filters
Least Square Regression
NA
ANU
Dense SIFT
Feed Forward Neural Networks
NA
IIITG
SURF
Multinomial Naïve Bayes
NA
Anonymous
SIFT, SURF, MLBP, PHOG
NA
Chi-square distance/ euclidean distance
Table 2. Features utilized by the participants’ algorithms in learning or non-learning based frameworks.
Evaluation Results by Organizers
Table 3: Equal error rates (EER) of the submitted algorithms per mobile device and lighting condition (daylight, office, and dim light) on VISIT 2
Figure 8: Bar graph showing the ranking of the submitted algorithms based on consolidated EER on VISIT 2. NTNU-1 outperforms all the other algorithms by obtaining an EER of 0.06% averaged over each device and lighting condition.
Table 4. EERs of the submitted algorithms across lighting conditions and different mobile devices on VISIT 2. The enrollment set consist of images acquired under normal office light and the verification set contains images from office, dim and day light environments.
Comparison of Results: VISIT 1 vs. VISIT 2 • VISIT 1 (Participants) and VISIT 2 (Organizers).
• Learning-based methods: When run on our end, VISIT 2 results were comparable to the ones reported by the participants on VISIT 1 (i.e., NTNU-1, NTNU-2, ANU and IIITG).
• Non-learning based method: However, the anonymous non-learning based submission showed large variation in the performance.
Conclusions • This is the first large-scale competition to compare ocular recognition schemes using mobile front facing captures. • Texture based analysis, when coupled with appropriate learning-based classifiers, showed high verification accuracy.
• The best result had an average EER of about 0.06% (NTNU-1). • No specific trend was observed in the performance of the algorithms across capturing devices and lighting conditions.
Contd.., • It can also be seen that, as expected, performance of all the algorithms degraded across lighting conditions. NTNU-1’s submission showed the least degradation in the performance. • Majority of the algorithms showed high degradation in cross lighting test (normal office light enrollment vs. dim light verification)
THANK YOU
