Covariate Analysis of Face Recognition Algorithms

1
Covariate Analysis of Face Recognition Algorithms

• Dr. J. Ross Beveridge
• Dr. Geof H. Givens
• Dr. Bruce Draper
• Mr. Yui Man Lui
• Colorado State University
• Dr. P. Jonathon Phillips
• National Institute of Standards and Technology

2
But Wait - Someone Said LIDAR
Work on LADAR (LIDAR) Recognition - 1997
3
Back to Faces - FRVT 2006

• NIST Sponsored Evaluations date back to the mid
  1990s
• Recent - FRVT 2006

4
Face Recognition Progress
Fig. 1 from FRVT 2006 Summary
Caveat - this is for well controlled imagery!
2006 - Falsely turn away 1/100 people, when only
admitting 1/1000 imposters.
5
Uncontrolled Lighting/Setting
6
FRVT 2006 Uncontrolled
Fig. 7 from FRVT 2006 Summary
Turn Away 20/100
2006 - Falsely turn away 10/100 to 40/100
people, when only admitting 1/1000 impostors.
7
Scope of the Study

• Algorithm - score fusion of 3 top performers.
• Imagery - Uncontrolled match to Controlled.
• Subset of FRVT 2006 Experiment 4
• 345 subjects and 110,514 match scores.

8
Scope - Covariates

• Performance Variable
• Verification Outcome, Success of Failure.
• False Accept Rate - FAR
• Properties of Environment
• Mugshot lighting, indoor uncontrolled, outdoor.
• Attributes of People
• Gender, Race, Age.
• Measurable Properties of Imagery
• Distance between Eyes.
• Face Region In Focus Measure (FRIFM).
• An edge-density measure by Eric Krotkov

Active Computer Vision by Cooperative Focus
and Stereo by Eric Krotkov.
9
Generalized Linear Mixed Model
Analysis is Mixed Effects Logistic
Regression with Repeated Measures on People.

• Let A and B be 2 covariates that might influence
  algorithm performance. For example, Agender
  (categorical) and BQuery-Eye-Distance
  (continuous).
• Let a index levels of A.
• Let j index the FAR setting, ?j
• Ypabj is
• 1 if Person p is verified correctly, 0 otherwise.
  
• Ypabj depends on
• person p, covariates A and B, and
• false alarm rate ?j.

Geof Givens, Statistics
10
Finding 2 Gender
11
Finding 4 Glasses
12
Face Region In Focus Measure

• FRIFM Sum of Sobel edge magnitude inside an
  ellipse bounding the face.

13
Face Region In Focus Measure
Low FRIFM examples
High FRIFM examples
14
Finding 5
Small
Medium
Large
15
Finding 5
Small
Medium
Large
Size of query image (distance between eyes)
16
Finding 5
Small
Medium
Large
Query environment
17
Finding 5
Small
Medium
Large
Boundary of observed data
18
Finding 5
Small
Medium
Large
Large PV range 0.90 ? 0.10
19
Finding 5
20

• Thank You

21
GLMM Model Continued
22
Subject Variation
The Mixed in Generalized Linear Mixed effect
Model.
This means
The outcomes, i. e. verification success/failure,
are uncorrelated when testing different people
but correlated when testing the same person under
different configurations.