Face Recognition-IACITS 2005

							        Face Recognition
Benchmarks, Caveats, Comparisons and
             Directions

       Gurumurthi V. Ramanan

       AU-KBC RESEARCH CENTRE
         MIT, ANNA UNIVERSITY
                CHENNAI
Face Recognition-The Problem



 Given a digitised image containing the face of a
person, extract the face region in the image and
identify or verify the person in the image, from a
database of face images.
    Face Recognition- From Face to
          Biometric Template
 The main steps involved in converting a face image into
   a biometric template (vector):
      •   Feature extraction, which includes face
          localisation, facial feature detection and extraction
      •   A pattern recognition methodology for classifying
          and comparing the data that has been extracted
 Each of these steps has its own success rate
      Face Recognition- From Face to
            Biometric Template
Feature extraction involves:
  • determination of the position of a single face (face
  localisation).
   Typical algorithm: Viola-Jones algorithm that locates the face
  using the statistics of skin texture. A literature survey in 1998
  mentions more than 150 face detection algorithms.
  • determination of the presence and location of features such
  as eyes, nose, lips (feature detection). All other features are
  usually located after first locating these.
  • a pattern recognition methodology that processes the data
  from the above two steps. This methodology will be used to
  classify and compare the faces.
  Face Recognition- Enrolment
Enrolment/ Training:

   • The system enrols an individual using digitised and
   good quality images containing the frontal face view.

   • This image is converted to a biometric template
   using pattern recognition techniques and stored in
   the database.

   • By a pattern recognition technique, we mean a
   mathematical model, which represents biometric
   templates as vectors in some high dimensional space
   with an inbuilt notion of similarity / distance.
    Face Recognition- Verification,
      Identification & Watch list
Verification (1-1):
      Are you who you say you are?
Identification (1-many):
      I know you are in my database. Can I find you?
 Watch list:
      Can I find you in my database? If I can, who are you?

Ranking: The system can also be made to report the top
matches from the database. These matches can be ordered by
rank. The system’s most likely candidate is assigned the rank 1.
A rank-5 success rate denotes the rate at which the correct
identity is within the top five matches.
Face Recognition- The Process
           Face Recognition-Rates
The verification (identification) rate is the rate at which
legitimate users are verified (identified).
 The false accept rate (FAR) is the rate at which impostors
are verified or identified by the system.
 The false reject rate (FRR) is the rate at which legitimate
users are wrongly verified (identified) as impostors by the
system.
The performance of a facial recognition system is measured
by the above rates. An ideal system would have a verification
rate of 100% and a false accept rate of 0%. Such systems do
not exist.
 Good systems balance this trade-off between the
verification rates and the false accept rates depending on the
application context and security needs.
      Face Recognition- Databases
One of the largest publicly available face database is the
FERET database. It consists of 14,051 eight-bit grayscale
images of human heads of 1072 subjects, with views
ranging from frontal to left and right profiles
The HCInt dataset used in the FRVT 2002 test consisted of
121,589 facial images of 37,437 persons. This is not a
publicly available database. This is supposed to be a part of a
much larger database of 6.8 million images collected from
Visa applicants at US Consular offices in a controlled
environment.
India does not have any such database.
    Face Recognition-Benchmarks
                    Verification Rates




Fig. 1. Verification performance is reported for all participants on
the HCInt visa dataset. Verification performance is reported at a
false accept rate of 1%. (Source: FRVT 2002)
    Face Recognition-Benchmarks
    Degradation in Verification Rates due
    to lighting conditions




Fig. 2. Verification performance is reported for five categories of frontal
facial images. Performance is reported for the best system and average of
the top three systems in each category. The verification rate is reported at
a false accept rate 1%. (Source: FRVT 2002)
    Face Recognition-Benchmarks
                  Identification Rates




Fig. 3. Identification performance for the three best systems on
the HCInt visa dataset. The database consisted of 37,437 persons.
Identification rates are reported for ranks 1, 10, and 50. (Source:
FRVT 2002)
   Face Recognition-Benchmarks
        Identification Rates due to age




Fig. 4. Identification performance is reported broken out by age of a
person. Each bin is labelled by the age range it contains (five year
intervals). Identification rate is the average for the top three systems.
Performance is on a database of 37,437. (Source: FRVT 2002)
       Face Recognition-Benchmarks
     Degradation in Identification Rates due to time




Fig. 5. Identification performance is reported broken out by elapsed
time between database and new image. Performance is reported in 60-
day intervals. The average rank one identification rate for the top three
systems is reported on a database of 37,437 persons. (Source: FRVT
2002)
 Face Recognition-Some Caveats
 Misperceptions about Commercial Products:
  • 100% match to any image at any angle
  • Instantly recognises any person
  • Database capabilities running to millions

 Reality (Research challenges):
  • Affected by lighting, angle,size of face,
    quality of captured and known image
  • Technology demonstrations done under lab /
    studio conditions
        Face Recognition- Factors
 Research done across Face Recognition (FR) algorithms
  indicate that they are susceptible to factors such as
   • Race, age, gender and time. These factors are not well
     quantified even in the research literature.
   • A recent study on the FERET database indicated
      - Wearing glasses makes people more recognisable
        contradicting previous study
      - Relative to the majority white faces, the Asian faces
        and African – American faces are significantly easier
        to recognise. (Total – 1072 subjects: White - 720,
        Asian - 143, African-American -121, other races- 88).
      - Different FR algorithms exhibit different behavior
        when confronted with factors such as gender.
         Face Recognition- Pilots
 In the US, State Departments of Motor Vehicles (DMV)
  are using software developed by Visionics and Polaroid, to
  prevent criminals from obtaining multiple licenses under
  different names.
 Computerised identity verification is in use by 37 of the 50
  American States. The Californian DMV database contains
  millions of images.
 The violence surrounding the Euro 2000 football games in
  the Netherlands was analyzed through CCTV cameras. FR
  systems recognized many individuals entering and leaving
  the country resulting in arrests by the Dutch authorities.
  http://news6.thdo.bbc.co.uk/1/hi/euro2000/teams/england/
  796242.stm.
          Face Recognition- Pilots
 Royal Canadian Mounted Police (RCMP) is using a face-
  scanning camera in the cell area of Pearson Airport to
  match people who are known criminals or terrorists.
 Anser also uses Visionics’ FaceIt system as part of their
  project with the National Center for Missing and Exploited
  Children to locate missing children on the Internet.
 In the July 2000 presidential elections in Mexico,
  Visionics’ FaceIt facial recognition tools were used to
  build a database of 8 million voters in efforts to eliminate
  voter fraud.
 Testing and internal market research is also big business
  within the biometric industry. The results of many of these
  pilots are confidential and some are sold as IP.
Face Recognition- Apprehensions
 In Super Bowl in Tampa Bay, Florida, FR system was used
  during the game and the week prior to it. The police
  claimed that they uncovered 19 people with criminal
  records in the crowd of over 100,000 at the Super Bowl.
  But American Civil Liberties Union (ACLU) claimed that
  ‘a study of the police logs showed that the system never
  correctly matched a face in its criminal database or
  resulted in any arrests’.
 Civil liberties groups fear a ‘function creep’ could occur
  i.e., biometrics could be introduced for one reason and then
  turned around and used for another different one.
Face Recognition vs. Fingerprints
 In 1893, the Home Ministry Office, UK, accepted that no
  two individuals have the same fingerprints leading to the
  development of Automatic Fingerprint Identification
  Systems (AFIS) in the 1960s. The current AFIS system at
  FBI consists of a large database of approximately 46
  million ‘ten prints’ and conducts, on an average,
  approximately 50,000 searches per day.
  Lesson: It takes time for wide acceptance of a new
  biometric.
 First modern systematic use of fingerprints seems to be in
  India in order to prevent the rich from paying the poor to
  serve in the prison in their place.
Face Recognition vs. Fingerprints
 Some aspersions on fingerprints :
   • stigma of criminality
   • fingerprint quality dependent on subject population and
     collection environment.
   • a 2004 fingerprint algorithms contest (NIST) revealed
     that fingerprint matching algorithms have false non-
     match error rate of 2%. This means a 100,000
     transactions/day (typical in a high throughput
     environment) would result in 2,000 false rejects/day.
   • lack of standards hampers inter-operability between
     proprietary but inexpensive systems.
Face Recognition vs. Fingerprints




 Verification rates of the top three FR systems vs. single fingerprint matcher
                 (Source: FRVT 2002 Evaluation report)
Face Recognition vs. Fingerprints
 Comparative performance:
  At false accept rates around 0.0l, verification
  performance is comparable. At false accept rates below
  0.01, fingerprint performance is better. At rates above
  0.01, the best face recognition systems perform better.
  For False accepts around 0.01, that face recognition
  performance is now comparable to large-scale
  fingerprint systems available in 1998. In 2005 FRVT,
  FR algorithms are expected to catch up.
         Face Recognition vs. Iris
 Iris has low error rates but these are not verifiable due to
  lack of publicly available databases. This hampers research
  as well as verifiability.
 Fragility in recent pilot studies - relatively high failure to
  enroll rates
   • An Iris pilot at a school reported 22% of unsuccessful
      transactions in a total of 9412 transactions over three
      months of which 16% was due to camera capture errors
      and 6% due to access attempts by unknown users.
 More research needs to be done before any meaningful
  comparisons can be made.
Face Recognition- Future Directions
  Face Recognition under outdoor lighting conditions
  Face Recognition for Surveillance with a watch-list
   size of at least 100 faces
  3D FR algorithms
  FR algorithms for victim identification
  FR algorithms for applications such as missing
   children
  FR algorithms that measure resemblance
  A deeper understanding of the manifold of faces
  Fusion of biometrics
Biometrics - The Grand Challenge
 For national id and using the body as a passport,
  these questions are the most challenging:

   • How to acquire repeatable and distinctive
     patterns from a broad population?

   • How to accurately and efficiently represent
     and recognise biometric patterns?
                                      (Anil Jain)
Thank you