ongoing face recognition vendor test (frvt) part 1 ... face recognition vendor test (frvt) part ......
TRANSCRIPT
-
NISTIR XXXX Draft
Ongoing Face RecognitionVendor Test (FRVT)Part 1: Verification
Patrick GrotherMei Ngan
Kayee HanaokaInformation Access Division
Information Technology Laboratory
This publication is available free of charge from:https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt-ongoing
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 1
DISCLAIMER
Specific hardware and software products identified in this report were used in order to perform the evalua-tions described in this document. In no case does identification of any commercial product, trade name, orvendor, imply recommendation or endorsement by the National Institute of Standards and Technology, nordoes it imply that the products and equipment identified are necessarily the best available for the purpose.
ABOUT THIS REPORT
This report is a draft NIST Interagency Report, and is open for comment. It documents the verification-trackof the ongoing Face Recognition Vendor Test. The report will be updated continuously as new algorithms areevaluated, as new datasets are added, and as new analyses are included. Comments and suggestions shouldbe directed to [email protected].
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 2
ContentsDISCLAIMER 1
1 CHANGELOG 6
2 METRICS 112.1 CORE ACCURACY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 DATASETS 123.1 CHILD EXPLOITATION IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123.2 VISA IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123.3 MUGSHOT IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123.4 SELFIE IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.5 WEBCAM IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.6 WILD IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4 RESULTS 144.1 TEST GOALS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144.2 TEST DESIGN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144.3 FAILURE TO ENROL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174.4 RECOGNITION ACCURACY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184.5 GENUINE DISTRIBUTION STABILITY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.5.1 EFFECT OF BIRTH PLACE ON THE GENUINE DISTRIBUTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334.5.2 EFFECT OF AGE ON GENUINE SUBJECTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.6 IMPOSTOR DISTRIBUTION STABILITY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374.6.1 EFFECT OF BIRTH PLACE ON THE IMPOSTOR DISTRIBUTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374.6.2 EFFECT OF AGE ON IMPOSTORS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
List of Tables1 ALGORITHM SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 FALSE NON-MATCH RATE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 FAILURE TO ENROL RATES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
List of Figures
1 PERFORMANCE SUMMARY: FNMR VS. TEMPLATE SIZE TRADEOFF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 PERFORMANCE SUMMARY: FNMR VS. TEMPLATE TIME TRADEOFF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 EXAMPLE IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
(A) VISA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14(B) MUGSHOT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14(C) WEBCAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14(D) SELFIE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14(E) WILD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4 ERROR TRADEOFF CHARACTERISTIC: VISA IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 ERROR TRADEOFF CHARACTERISTIC: VISA IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 ERROR TRADEOFF CHARACTERISTIC: MUGSHOT IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 ERROR TRADEOFF CHARACTERISTIC: SELFIE IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 ERROR TRADEOFF CHARACTERISTIC: SELFIE IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 ERROR TRADEOFF CHARACTERISTIC: WILD IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2410 ERROR TRADEOFF CHARACTERISTICS: CHILD EXPLOITATION IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 3
11 ERROR TRADEOFF CHARACTERISTICS: CHILD EXPLOITATION IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2612 SEX AND RACE EFFECTS: MUGSHOT IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2713 SEX EFFECTS: VISA IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2814 ERROR TRADEOFF CHARACTERISTIC: WILD IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2915 FALSE MATCH RATE CALIBRATION: VISA IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3016 FALSE MATCH RATE CONCENTRATION: VISA IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3117 FALSE MATCH RATE CALIBRATION: MUGSHOT IMAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3218 EFFECT OF COUNTRY OF BIRTH ON FNMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3419 EFFECT OF SUBJECT AGE ON FNMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3620 IMPOSTOR DISTRIBUTION SHIFTS FOR SELECT COUNTRY PAIRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3921 ALGORITHM 3DIVI-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4022 ALGORITHM AYONIX-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4123 ALGORITHM CYBEREXTRUDER-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4224 ALGORITHM DERMALOG-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4325 ALGORITHM DERMALOG-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4426 ALGORITHM DERMALOG-003 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4527 ALGORITHM DIGITALBARRIERS-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4628 ALGORITHM DIGITALBARRIERS-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4729 ALGORITHM ID3-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4830 ALGORITHM ID3-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4931 ALGORITHM INNOVATRICS-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5032 ALGORITHM INNOVATRICS-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5133 ALGORITHM ISITYOU-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5234 ALGORITHM ITMO-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5335 ALGORITHM ITMO-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5436 ALGORITHM MORPHO-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5537 ALGORITHM NEUROTECHNOLOGY-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5638 ALGORITHM NEUROTECHNOLOGY-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5739 ALGORITHM NTECHLAB-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5840 ALGORITHM NTECHLAB-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5941 ALGORITHM RANKONE-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6042 ALGORITHM RANKONE-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6143 ALGORITHM RANKONE-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6244 ALGORITHM SAMTECH-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6345 ALGORITHM TONGYITRANS-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6446 ALGORITHM TONGYITRANS-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6547 ALGORITHM TUPEL-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6648 ALGORITHM VCOG-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6749 ALGORITHM VCOG-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6850 ALGORITHM VIGILANTSOLUTIONS-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6951 ALGORITHM VIGILANTSOLUTIONS-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7052 ALGORITHM VISIONLABS-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7153 ALGORITHM VOCORD-001 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7254 ALGORITHM VOCORD-002 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7355 ALGORITHM YITU-000 CROSS REGION FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7456 ALGORITHM 3DIVI-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7557 ALGORITHM AYONIX-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7658 ALGORITHM CYBEREXTRUDER-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7759 ALGORITHM DERMALOG-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7860 ALGORITHM DERMALOG-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7961 ALGORITHM DERMALOG-003 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8062 ALGORITHM DIGITALBARRIERS-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8163 ALGORITHM DIGITALBARRIERS-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8264 ALGORITHM ID3-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8365 ALGORITHM ID3-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8466 ALGORITHM INNOVATRICS-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8567 ALGORITHM INNOVATRICS-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 4
68 ALGORITHM ISITYOU-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8769 ALGORITHM ITMO-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8870 ALGORITHM ITMO-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8971 ALGORITHM MORPHO-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9072 ALGORITHM NEUROTECHNOLOGY-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9173 ALGORITHM NEUROTECHNOLOGY-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9274 ALGORITHM NTECHLAB-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9375 ALGORITHM NTECHLAB-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9476 ALGORITHM RANKONE-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9577 ALGORITHM RANKONE-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9678 ALGORITHM RANKONE-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9779 ALGORITHM SAMTECH-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9880 ALGORITHM TONGYITRANS-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9981 ALGORITHM TONGYITRANS-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10082 ALGORITHM TUPEL-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10183 ALGORITHM VCOG-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10284 ALGORITHM VCOG-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10385 ALGORITHM VIGILANTSOLUTIONS-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10486 ALGORITHM VIGILANTSOLUTIONS-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10587 ALGORITHM VISIONLABS-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10688 ALGORITHM VOCORD-001 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10789 ALGORITHM VOCORD-002 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10890 ALGORITHM YITU-000 CROSS COUNTRY FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10991 IMPOSTOR COUNTS FOR CROSS COUNTRY FMR CALCULATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11092 ALGORITHM 3DIVI-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11293 ALGORITHM AYONIX-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11394 ALGORITHM CYBEREXTRUDER-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11495 ALGORITHM DERMALOG-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11596 ALGORITHM DERMALOG-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11697 ALGORITHM DERMALOG-003 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11798 ALGORITHM DIGITALBARRIERS-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11899 ALGORITHM DIGITALBARRIERS-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119100 ALGORITHM ID3-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120101 ALGORITHM ID3-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121102 ALGORITHM INNOVATRICS-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122103 ALGORITHM INNOVATRICS-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123104 ALGORITHM ISITYOU-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124105 ALGORITHM ITMO-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125106 ALGORITHM ITMO-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126107 ALGORITHM MORPHO-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127108 ALGORITHM NEUROTECHNOLOGY-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128109 ALGORITHM NEUROTECHNOLOGY-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129110 ALGORITHM NTECHLAB-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130111 ALGORITHM NTECHLAB-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131112 ALGORITHM RANKONE-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132113 ALGORITHM RANKONE-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133114 ALGORITHM RANKONE-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134115 ALGORITHM SAMTECH-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135116 ALGORITHM TONGYITRANS-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136117 ALGORITHM TONGYITRANS-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137118 ALGORITHM TUPEL-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138119 ALGORITHM VCOG-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139120 ALGORITHM VCOG-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140121 ALGORITHM VIGILANTSOLUTIONS-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141122 ALGORITHM VIGILANTSOLUTIONS-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142123 ALGORITHM VISIONLABS-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143124 ALGORITHM VOCORD-001 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 5
125 ALGORITHM VOCORD-002 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145126 ALGORITHM YITU-000 CROSS AGE FMR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 6
1 Changelog
2017-08-22
. Added results for three additional algorithms, rankone-002, neurotechnology-001, and itmo-002.
. The algorithms dermalog-001 and rank-001 have been retired - FRVT lists only two algorithms per organization
. The algorithm tupel-001 has been retired as it is not operable on all datasets
. Clarified the tradeoff Figures 1 and 2 plot only genuine comparison durations.
. Corrected image type label in section 3.5
2017-08-07
. Added results for 5 new algorithms
. Added Figure 3 giving simulated example images.
. Added Figure 1 showing an alternative view of the same tradeoff data in Figure 2
. Added Figure 5 showing accuracy on visa images just for low FMR.
. Added Figure 20 showing impostor distribution shifts from certain country pairs. Section 4.6.1 in this and priorreports documents high false match rates for individuals from certain countries. That effect, however, is often notconfined to anomalously high impostor scores in the tails of the distribution, but arises from systematic shifts ofthe whole distribution. These shifts sometimes reach 2.
2017-07-29
. Added results for 8 new algorithms
. Added results for a child-exploitation dataset
. Added Table 2 a standalone tabulation of false non-match rates
. We have received additional CPU algorithms - Results should appear August 4, 2017
. We have received additional GPU algorithms - Results to appear as computational resources are released from theFace Recognition Prize Challenge
2017-06-19
. Added five new algorithms, three of which remain in-process
. Added results for a wild dataset of images similar to non-cooperative photojournalism images
. Added Table 3 a standalone tabulation of failure to enrol rates
. Added Fig. 2 showing tradeoff between FNMR, template size, template generation time, and match duration.
. Added Fig. 16 showing how FMR is concentrated in certain images.
. Restated cross-region false match rates at nominal FMR = 0.0001 instead of 0.001
. Improved DET legends.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
https://www.nist.gov/programs-projects/face-recognition-prize-challenge
-
FR
VT
-FA
CE
RE
CO
GN
ITIO
NV
EN
DO
RT
EST
-V
ER
IFICA
TIO
N7
Developer Short Seq. Validation Config1 Template GPU Comparison Time (ns)3
Name Name Num. Date Data (KB) Size (B) Time (ms)2 Genuine Impostor
1 3DiVi 3divi 000 2017-03-16 169360 8512 0 13285 52 Yes 1378 20 2375 192 Ayonix ayonix 000 2017-06-22 58505 111036 0 118 2 No 2621 23 3620 263 Cyberextruder cyberex 001 2017-08-02 121211 6256 0 28893 25 No 61083 16 81079 194 Dermalog dermalog 002 2017-02-22 0 131043 0 381 0 No 1822169 114 2122105 1465 Dermalog dermalog 003 2017-07-10 0 121043 0 9121 22 No 1922957 93 2222808 1316 Digital Barriers barriers 000 2017-05-31 157794 172056 0 5104 0 No 1413232 166 1713226 1467 Digital Barriers barriers 001 2017-07-20 236934 182056 0 14294 1 No 1312311 164 1512347 1978 ID3 Technology id3 001 2017-08-04 225574 10520 0 10238 19 No 51058 32 71049 289 ID3 Technology id3 002 2017-08-04 225574 9520 0 18482 34 No 71100 59 61048 3210 Innovatrics innova 000 2017-07-25 0 3146 0 19578 5 No 104964 63 114665 26211 Innovatrics innova 001 2017-07-25 0 7284 0 23645 5 No 115506 131 124975 30812 Is It You isityou 000 2017-06-26 48010 2919200 0 8113 5 No 30237517 1318 30237374 127913 ITMO University itmo 001 2017-06-12 1923215 3137997 0 32959 18 No 2028901 1492 2327517 18614 ITMO University itmo 002 2017-08-07 1923215 254162 0 21611 17 No 127423 96 137451 9415 Morpho morpho 000 2017-07-11 100806 1116 0 7109 1 No 4993 31 51000 3416 Neurotechnology neurotech 000 2017-03-22 62129 287148 0 22611 48 No 2874288 2194 2872879 264017 Neurotechnology neurotech 001 2017-08-07 280771 212718 0 26881 46 No 2769356 684 2769140 57918 N-Tech Lab ntech 000 2017-03-13 191530 222906 1 12278 13 No 2230787 142 2430846 7719 N-Tech Lab ntech 001 2017-05-10 691296 276744 1 20587 11 No 2567692 833 2667486 24420 Rank One Computing rankone 000 2017-03-21 0 2144 0 482 9 No 2439932 468 148722 17121 Rank One Computing rankone 002 2017-08-18 0 5224 0 275 1 No 2669113 3802 1369 2622 Samtech InfoNet Limited samtech 000 2017-05-02 109774 162056 0 11262 2 No 94550 26 104541 2823 TongYi Transportation Technology tongyi 001 2017-04-01 625339 202058 0 15310 20 No 1617769 74 1917750 6324 TongYi Transportation Technology tongyi 002 2017-07-15 625336 192058 0 16356 35 No 2129816 281 2017799 12725 VCognition vcog 001 2017-03-28 86103 234126 0 6108 17 Yes 1516320 197 1816426 42526 VCognition vcog 002 2017-06-12 3229434 3261504 5 17357 25 No 31296154 3077 31296436 418327 Vigilant Solutions vigilant 000 2017-03-30 352218 3031540 0 27884 23 No 1718201 94 1613030 8328 Vigilant Solutions vigilant 001 2017-06-13 344685 151544 0 30921 2 No 3644 13 4649 1629 VisionLabs visionlabs 001 2017-06-12 343661 4204 0 31943 8 No 81395 45 91148 5330 Vocord vocord 001 2017-04-21 616989 266194 0 29908 16 No 321094730 64282 321107193 6652331 Vocord vocord 002 2017-06-07 918292 141330 0 25782 36 Yes 2983063 517 2983072 71432 Shanghai Yitu Technology yitu 000 2017-05-23 2211068 244130 0 24672 2 No 2335352 114 2537848 1773
Notes1 The size of configuration data does not capture static data included in the libraries. We do not include the size of the libraries because
some algorithms include common ancilliary libraries for image processing (e.g. openCV) or numerical computation (e.g. blas).2 The median template creation times are measured on Intel RXeon RCPU E5-2630 v4 @ 2.20GHz processors or, in the case of GPU-enabled
implementations, NVidia Tesla K40.3 The median comparison durations, in nanoseconds, are estimated using std::chrono::high resolution clock which on the machine in (2)
counts clock ticks of duration 1ns. Precision is somewhat worse than that however. The value is the median absolute deviation times1.48 for Normal consistency.
Table 1: Summary of algorithms and properties included in this report. The red superscripts give ranking for the quantity in that column.
2017/08/
2513:00:41
FNM
R(T)
Falsenon-m
atchrate
FMR
(T)False
match
rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 8
FALSE NON-MATCH RATE (FNMR)Algorithm CONSTRAINED, COOPERATIVE LESS CONSTRAINED, NON-COOPERATIVEName VISA VISA MUGSHOT WEBCAM SELFIE WILD CHILD EXP
FMR 1E-06 0.0001 0.0001 0.0001 0.0001 0.0001 0.01
1 3divi-000 0.133 7 0.029 6 0.037 9 0.002 7 0.055 14 0.547 4 0.553 32 ayonix-000 0.487 27 0.230 28 0.309 28 0.172 27 0.360 28 0.807 20 0.843 163 cyberextruder-001 0.255 18 0.076 20 0.165 22 0.029 19 0.144 24 0.853 23 - 314 dermalog-002 0.315 22 0.122 23 0.241 26 0.109 26 0.179 26 0.709 12 0.985 21
5 dermalog-003 0.280 20 0.112 22 0.202 25 0.041 23 0.115 23 0.693 9 0.845 186 digitalbarriers-000 0.463 26 0.161 25 0.184 24 0.045 24 0.170 25 0.741 17 0.771 147 digitalbarriers-001 0.502 28 0.155 24 0.041 12 0.029 20 0.115 22 0.678 8 - 258 id3-001 0.250 17 0.063 18 0.036 8 0.002 8 0.040 10 0.765 18 - 32
9 id3-002 0.239 16 0.057 17 0.032 7 0.003 10 0.037 8 0.810 21 - 2310 innovatrics-000 0.191 12 0.034 10 0.046 17 0.001 4 0.040 9 0.720 13 - 2411 innovatrics-001 0.183 11 0.034 9 0.043 13 0.001 5 0.043 11 0.643 6 - 2812 isityou-000 0.703 30 0.414 30 0.680 30 0.690 31 - 32 1.000 31 - 22
13 itmo-001 0.441 24 0.174 27 0.256 27 0.069 25 0.199 27 0.932 28 - 2714 itmo-002 0.287 21 0.050 15 0.047 18 0.036 21 0.069 20 0.892 25 - 2615 morpho-000 0.134 8 0.026 4 0.028 4 0.007 13 0.012 2 0.893 26 0.846 1916 neurotechnology-000 0.237 15 0.052 16 0.062 19 0.005 12 0.052 12 0.943 29 0.845 17
17 neurotechnology-001 0.222 14 0.044 13 0.046 16 0.001 6 0.017 7 0.817 22 - 2918 ntechlab-000 0.086 6 0.027 5 0.044 14 0.003 9 0.014 5 0.367 2 0.533 219 ntechlab-001 0.083 5 0.025 3 0.030 5 0.003 11 0.014 4 0.319 1 0.472 120 rankone-000 0.276 19 0.071 19 0.177 23 0.021 17 0.092 21 0.723 14 0.787 15
21 rankone-002 0.217 13 0.049 14 0.032 6 0.000 2 0.052 13 0.705 11 - 3022 samtech-000 0.443 25 0.161 26 0.044 15 0.021 18 0.063 18 0.878 24 0.765 1323 tongyitrans-001 0.072 4 0.038 12 0.041 11 0.009 14 0.063 16 0.704 10 0.743 924 tongyitrans-002 0.066 3 0.030 7 0.039 10 0.010 15 0.063 17 0.725 15 0.746 10
25 vcog-001 0.892 31 0.409 29 - 32 0.302 28 0.427 29 - 32 0.686 626 vcog-002 0.903 32 0.504 32 0.692 31 0.559 30 0.666 31 0.778 19 0.752 1127 vigilantsolutions-000 0.688 29 0.415 31 0.595 29 0.401 29 0.643 30 0.915 27 0.894 2028 vigilantsolutions-001 0.348 23 0.105 21 0.101 21 0.016 16 0.061 15 0.729 16 0.730 8
29 visionlabs-001 0.180 10 0.030 8 0.024 3 0.001 3 0.014 6 0.591 5 0.561 430 vocord-001 0.141 9 0.035 11 0.063 20 0.036 22 0.069 19 0.654 7 0.695 731 vocord-002 0.034 2 0.013 1 0.019 2 - 32 0.012 3 0.948 30 0.762 1232 yitu-000 0.033 1 0.021 2 0.017 1 0.000 1 0.012 1 0.431 3 0.586 5
Table 2: FNMR is the proportion of mated comparisons below a threshold set to achieve the FMR given in the header on the fourthrow. FMR is the proportion of impostor comparisons at or above that threshold. Note that the webcam and selfie values apply toimages collected on the same day, and that will often yield optimisticlly low FNMR values.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 9
3div
i0
ayon
ix0
cybe
rex1
derm
alo2
derm
alo3
digi
tal0
digi
tal1
id31
id32
inno
vat0
inno
vat1
isity
ou0
itmo1
itmo2
mor
pho0
neur
ote0
neur
ote1
ntec
hla0
ntec
hla1
rank
one0
rank
one2
sam
tech
0
tong
yit1
tong
yit2
vigi
lan0
vigi
lan1
visi
onl1
voco
rd1
voco
rd2
yitu
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
0.20
0.30
0.40
0.50
0.60
0.70
100
200
300
400
500
600
700
800
1000
2000
3000
4000
5000
7000
9000
3000
050
000
Tem
plat
e si
ze (
byte
s)
FNMR at FMR = 0.0001
3456lo
g10
(mat
ch_n
ano
sec)
log
10(t
emp
late
_mill
isec
)
1.5
2.0
2.5
Figu
re1:
The
poin
tssh
owfa
lse
non-
mat
chra
tes
(FN
MR
)ve
rsus
the
dura
tion
ofth
ete
mpl
ate
gene
rati
onop
erat
ion.
FNM
Ris
the
geom
etri
cm
ean
ofFN
MR
valu
esfo
rvi
saan
dm
ugsh
otim
ages
(fro
mFi
gs.4
and
6)at
afa
lse
mat
chra
te(F
MR
)of0
.000
1.Te
mpl
ate
gene
rati
onti
me
isa
med
ian
esti
mat
edov
er64
0x
480
pixe
lpor
trai
ts.T
hesi
zeof
the
poin
tsen
code
ste
mpl
ate
gene
rati
onti
me
-w
hich
span
son
eor
der
ofm
agni
tude
.Th
eco
lor
ofth
epo
ints
enco
des
one-
to-o
nege
nuin
ete
mpl
ate
com
pari
son
dura
tion
-whi
chsp
anth
ree
orde
rsof
mag
nitu
de.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 10
3div
i0
cybe
rex1
derm
alo2
derm
alo3
digi
tal0
digi
tal1
id31
id32
inno
vat0
inno
vat1
isity
ou0
itmo1
itmo2
mor
pho0
neur
ote0
neur
ote1
ntec
hla0
ntec
hla1
rank
one0
rank
one2
sam
tech
0
tong
yit1
tong
yit2
vcog
2
vigi
lan0
vigi
lan1
visi
onl1
voco
rd1
voco
rd2
yitu
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
0.20
0.30
0.40
0.50
0.60
0.70
5060
7080
9010
020
030
040
050
060
070
080
090
010
00Te
mpl
ate
gene
ratio
n du
ratio
n (m
illis
econ
ds)
FNMR at FMR = 0.0001
3456lo
g10
(mat
ch_n
ano
sec)
log
10(t
emp
late
_byt
es)
3 4
Figu
re2:
The
poin
tssh
owfa
lse
non-
mat
chra
tes
(FN
MR
)ve
rsus
the
dura
tion
ofth
ete
mpl
ate
gene
rati
onop
erat
ion.
FNM
Ris
the
geom
etri
cm
ean
ofFN
MR
valu
esfo
rvi
saan
dm
ugsh
otim
ages
(fro
mFi
gs.4
and
6)at
afa
lse
mat
chra
te(F
MR
)of0
.000
1.Te
mpl
ate
gene
rati
onti
me
isa
med
ian
esti
mat
edov
er64
0x
480
pixe
lpor
trai
ts.T
hesi
zeof
the
poin
tsen
code
ste
mpl
ate
size
-w
hich
span
two
orde
rsof
mag
nitu
de.
The
colo
rof
the
poin
tsen
code
son
e-to
-one
genu
ine
tem
plat
eco
mpa
riso
ndu
rati
on-
whi
chsp
anth
ree
orde
rsof
mag
nitu
de.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 11
2 Metrics
2.1 Core accuracy
Given a vector of N genuine scores, u, the false non-match rate (FNMR) is computed as the proportion below somethreshold, T:
FNMR(T ) = 1 1N
Ni
H(ui T ) (1)
where H(x) is the unit step function, and H(0) taken to be 1.
Similarly, given a vector of N impostor scores, v, the false match rate (FMR) is computed as the proportion above T:
FMR(T ) =1
N
Ni
H(vi T ) (2)
The threshold, T, can take on any value. We typically generate a set of thresholds from quantiles of the observed impostorscores, v, as follows. Given some interesting false match rate range, [FMRL,FMRU ], we form a vector of K thresholdscorresponding to FMR measurements evenly spaced on a logarithmic scale
Tk = Qv(1 FMRk) (3)
where Q is the quantile function, and FMRk comes from
log10 FMRk = log10 FMRL +k
K[log10 FMRU log10 FMRL] (4)
Error tradeoff characteristics are plots of FNMR(T) vs. FMR(T). These are plotted with FMRU 1 and FMRL as lowas is sustained by the number of impostor comparisons, N. This is somewhat higher than the rule of three limit 3/Nbecause samples are not independent, due to re-use of images.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 12
3 Datasets
3.1 Child exploitation images
. The number of images is O(104).
. The number of subjects is O(103).
. The number of subjects with two images O(103).
. The images are operational. They are taken from ongoing investigations of child exploitation crimes. The imagesare arbitrarily unconstrained. Pose varies considerably around all three axes, including subject lying down. Res-olution varies very widely. Faces can be occluded by other objects, including hair and hands. Lighting varies,although the images are intended for human viewing. Mis-focus is rare. Images are given to the algorithm withoutany cropping; faces may occupy widely varying areas.
. The images are usually large from contemporary cameras. The mean interocular distance (IOD) is 70 pixels.
. The images are of subjects from several countries, due to the global production of this imagery.
. The images are of children, from infancy to late adolescence.
. All of the images are live capture, none are scanned. Many have been cropped.
. When these images are input to the algorithm, they are labelled as being of type EXPLOITATION - see Table 4 ofthe FRVT API.
3.2 Visa images
. The number of images is O(105).
. The number of subjects is O(105).
. The number of subjects with two images O(104).
. The images have geometry in reasonable conformance with the ISO/IEC 19794-5 Full Frontal image type. Pose isgenerally excellent.
. The images are of size 252x300 pixels. The mean interocular distance (IOD) is 69 pixels.
. The images are of subjects from greater than 100 countries, with significant imbalance due to visa issuance patterns.
. The images are of subjects of all ages, including children, again with imbalance due to visa issuance demand.
. Many of the images are live capture. A substantial number of the images are photographs of paper photographs.
. When these images are input to the algorithm, they are labelled as being of type ISO - see Table 4 of the FRVTAPI.
3.3 Mugshot images
. The number of images is O(106).
. The number of subjects is O(105).
. The number of subjects with two images O(105).
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 13
. The images have geometry in reasonable conformance with the ISO/IEC 19794-5 Full Frontal image type.
. The images are of variable sizes. The median IOD is 104 pixels. The mean IOD is 123 pixels.
. The images are of subjects from the United States.
. The images are of adults.
. The images are all live capture.
. When these images are input to the algorithm, they are labelled as being of type mugshot - see Table 4 of theFRVT API.
3.4 Selfie images
. The number of images is below 500.
. The number of subjects is below 500.
. All subjects have a selfie image, and a portrait image.
. The portrait images are in reasonable conformance with the ISO/IEC 19794-5 Full Frontal image type.
. The selfie images vary: taken with camera above and below eye level, with one hand or two hands. Pitch anglesvary more than yaw angles, which are frontal. Some perspective distortion is evident.
. The images have mean IOD of 140 pixels.
. The images are of subjects from the United States.
. The images are of adults.
. The images are all live capture.
. When these images are input to the algorithm, they are labelled as being of type WILD - see Table 4 of the FRVTAPI.
3.5 Webcam images
. The number of images is below 1500.
. The number of subjects is below 1500.
. All subjects have a webcam image, and a portrait image.
. The portrait images are in reasonable conformance with the ISO/IEC 19794-5 Full Frontal image type.
. The webcam images are taken with camera at a typical head height, with mild pitch angles, low yaw angles, butsome variation in range, such that low perspective distortion is sometimes evident.
. The images have mean IOD of 68 pixels (sd=12).
. The images are of subjects from the United States.
. The images are of adults.
. The images are all live capture.
. When these images are input to the algorithm, they are labelled as being of type MUGSHOT - see Table 4 of theFRVT API.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 14
(a) Visa (b) Mugshot (c) Webcam (d) Selfie (e) Wild
Figure 3: The figure gives simulated samples of image types used in this report.
3.6 Wild images
. The number of images is O(105).
. The number of subjects is O(103).
. The number of subjects with two images O(103).
. The images include many photojournalism-style images. Images are given to the algorithm using a variable butgenerally tight crop of the head. Resolution varies very widely. The images are very unconstrained, with wide yawand pitch pose variation. Faces can be occluded, including hair and hands.
. The images are of adults.
. All of the images are live capture, none are scanned.
. When these images are input to the algorithm, they are labelled as being of type WILD - see Table 4 of the FRVTAPI.
4 Results
4.1 Test goals
. To state overall accuracy.
. To compare algorithms.
4.2 Test design
Method: For visa images:
. The comparisons are of visa photos against visa photos.
. The number of genuine comparisons is O(104).
. The number of impostor comparisons is O(1010).
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 15
. The comparisons are fully zero-effort, meaning impostors are paired without attention to sex, age or other covari-ates. However, later analysis is conducted on subsets.
. The number of persons is O(105).
. The number of images used to make 1 template is 1.
. The number of templates used to make each comparison score is two corresponding to simple one-to-one verifica-tion.
For mugshot images:
. The comparisons are of mugshot photos against mugshot photos.
. The number of genuine comparisons is O(105).
. The number of impostor comparisons is O(107).
. The comparisons are fully zero-effort, meaning impostors are paired without attention to sex, age or other covari-ates.
. The number of persons is O(106).
. The number of images used to make 1 template is 1.
. The number of templates used to make each comparison score is two corresponding to simple one-to-one verifica-tion.
For selfie images:
. The comparisons are of selfie photos against portrait photos.
. The number of genuine comparisons is O(102).
. The number of impostor comparisons is O(108) selfies are compared with portraits of O(106) other subjects.
. The comparisons are fully zero-effort, meaning impostors are paired without attention to sex, age or other covari-ates.
. The number of persons is O(106).
. The number of images used to make 1 template is 1.
. The number of templates used to make each comparison score is two corresponding to simple one-to-one verifica-tion.
For webcam images:
. The comparisons are of webcam photos against portrait photos.
. The number of genuine comparisons is O(103).
. The number of impostor comparisons is O(109) webcams are compared with portraits of O(106) other subjects.
. The comparisons are fully zero-effort, meaning impostors are paired without attention to sex, age or other covari-ates.
. The number of persons is O(106).
. The number of images used to make 1 template is 1.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 16
. The number of templates used to make each comparison score is two corresponding to simple one-to-one verifica-tion.
For child exploitation images:
. The comparisons are of unconstrained child exploitation photos against others of the same type.
. The number of genuine comparisons is O(104).
. The number of impostor comparisons is O(107).
. The comparisons are fully zero-effort, meaning impostors are paired without attention to sex, age or other covari-ates.
. The number of persons is O(103).
. The number of images used to make 1 template is 1.
. The number of templates used to make each comparison score is two corresponding to simple one-to-one verifica-tion.
. We produce two performance statements. First, is a DET as used for visa and mugshot images. The second isa cumulative match characteristic (CMC) summarizing a simulated one-to-many search process. This is done asfollows.
We regard M enrollment templates as items in a gallery.
These M templates come from M > N individuals, because multiple images of a subject are present in thegallery under separate identifiers.
We regard the verification templates as search templates.
For each search we compute the rank of the highest scoring mate.
This process should properly be conducted with a 1:N algorithm, such as those tested in NIST IR 8009. Weuse the 1:1 algorithms in a simulated 1:N mode here to a) better reflect what a child exploitation analyst does,and b) to do show algorithm efficacy is better than that revealed in the verification DETs.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 17
4.3 Failure to enrol
Algorithm Failure to Enrol Rate1
Name CHILD-EXPLOIT MUGSHOT SELFIES VISA WEBCAM WILD
3divi-000 0.2019 12 0.0019 18 0.0202 18 0.0008 12 0.0020 15 0.2070 19ayonix-000 0.0000 1 0.0109 31 0.0751 29 0.0137 35 0.0109 21 0.0000 2cyberextruder-001 - 32 0.0036 22 0.0376 20 0.0029 24 0.0205 26 0.5730 32dermalog-002 0.9109 26 0.0045 26 0.0954 32 0.0013 15 0.0471 30 0.3979 26
dermalog-003 0.0434 6 0.0007 7 0.0000 1 0.0025 23 0.0007 6 0.0701 11digitalbarriers-000 0.5469 23 0.0043 24 0.0925 30 0.0019 21 0.0184 25 0.5170 31digitalbarriers-001 - 32 0.0044 25 0.0925 31 0.0018 20 0.0232 27 0.4593 30id3-001 - 32 0.0043 23 0.0260 19 0.0043 32 0.0014 12 0.2090 20
id3-002 - 32 0.0030 21 0.0202 17 0.0032 26 0.0020 14 0.1496 14innovatrics-000 - 32 0.0013 11 0.0087 13 0.0004 8 0.0027 16 0.3337 24innovatrics-001 - 32 0.0013 12 0.0087 15 0.0004 9 0.0027 17 0.3337 25isityou-000 0.4714 21 0.0022 20 0.0665 27 0.0010 13 0.0116 23 0.4586 29
itmo-001 - 32 0.0114 32 0.0694 28 0.0047 33 0.0546 32 0.7410 36itmo-002 - 32 0.0068 29 0.0636 26 0.0029 25 0.0498 31 0.6985 34morpho-000 0.0000 2 0.0000 1 0.0000 4 0.0000 2 0.0000 4 0.0000 5neurotechnology-000 0.0000 4 0.0000 3 0.0000 5 0.0000 3 0.0000 5 0.0163 8
neurotechnology-001 - 32 0.0000 2 0.0000 2 0.0000 1 0.0000 1 0.0163 7ntechlab-000 0.2496 15 0.0015 16 0.0058 11 0.0016 18 0.0007 7 0.1099 13ntechlab-001 0.0926 10 0.0009 8 0.0029 7 0.0005 10 0.0007 8 0.0584 10rankone-000 0.0187 5 0.0005 5 0.0000 6 0.0003 6 0.0007 10 0.2349 21
rankone-002 - 32 0.0001 4 0.0000 3 0.0000 4 0.0000 3 0.0855 12samtech-000 0.5474 24 0.0052 27 0.0491 24 0.0042 31 0.0252 28 0.7023 35tongyitrans-001 0.0000 3 0.0068 28 0.0462 22 0.0040 29 0.0055 19 0.0000 3tongyitrans-002 0.3609 19 0.0078 30 0.0462 23 0.0040 30 0.0055 20 0.0000 4
vcog-001 0.1579 11 - 32 0.0058 10 0.0018 19 0.0000 2 - 32vcog-002 0.2209 13 0.0021 19 0.0087 14 0.0019 22 0.0007 9 0.1672 16vigilantsolutions-000 0.5580 25 0.0018 17 0.0462 21 0.0007 11 0.0109 22 0.5927 33vigilantsolutions-001 0.3585 18 0.0010 10 0.0116 16 0.0004 7 0.0048 18 0.3262 23
visionlabs-001 0.2699 16 0.0014 13 0.0058 12 0.0014 17 0.0020 13 0.1803 17vocord-001 0.4732 22 0.0158 33 0.0520 25 0.0038 28 0.0348 29 0.4494 28vocord-002 0.3782 20 0.0015 14 0.0029 9 0.0037 27 0.0171 24 0.1992 18yitu-000 0.3475 17 0.0015 15 0.0029 8 0.0013 16 0.0014 11 0.1591 15
Table 3: FTE is the proportion of failed template generation attempts. Failures can occur because the software throws an exception,or because the software electively refuses to process the input image. This would typically occur if a face is not detected. FTE ismeasured as the number of function calls that give a non-zero error code, OR that give a small template. This is defined as onewhose size is less than 0.3 times the median template size. This second rule is needed because some algorithms incorrectly fail toreturn a non-zero error code when template generation fails.1The effects of FTE are included in the accuracy results of this report by regarding any template comparison involving a failed templateenrollment to produce a low similarity score. Thus higher FTE results in higher FNMR.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 18
4.4 Recognition accuracy
Core algorithm accuracy is stated via:
. Cooperative subjects
The summary table of Figure 2;
The visa image DETs of Figures 4 and 5;
The mugshot DETs of Figure 6 ;
The selfie-portrait DETs of Figure 7;
The webcam-portrait DETs of Figure 8;
. Cooperative subjects
The photojournalism DET of Figure 9
The child-exploitation DET of Figure 10;
The child-exploitation CMC of Figure 11.
Figure 15 shows dependence of false match rate on algorithm score threshold. This allows a deployer to set a thresholdto target a particular false match rate appropriate to the security objectives of the application.
Figure 17 likewise shows FMR(T) but for mugshots, and specially four subsets of the population.
Note that in both the mugshot and visa sets false match rates vary with the ethnicity, age, and sex, of the enrollee andimpostor - see section 4.6.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 19
isit0
rank
0
rank
2ne
ur0
ntec
0
ntec
1
itmo1
ayon
0
vcog
1
vcog
2
vigi
0
vigi
1di
gi0
itmo2
derm
2
derm
3
tong
1
tong
2
sam
t0
yitu
0
visi
1
voco
1
voco
2
3div
0
mor
p0
inno
0
0.00
3
0.01
0
0.03
0
0.10
0
0.30
0 1e
071e
06
1e
051e
04
1e
031e
02
1e
011e
+00
Fals
e m
atch
rat
e (F
MR
)
False nonmatch rate (FNMR)
Dat
aset
: V
ISA
FN
MR
@ F
MR
=0.0
001
and
Alg
ori
thm
0.50
2 v
cog_
002
0.41
5 v
igila
ntso
lutio
ns_0
00
0.41
3 i
sity
ou_0
00
0.40
7 v
cog_
001
0.22
9 a
yoni
x_00
0
0.17
3 i
tmo_
001
0.16
0 d
igita
lbar
riers
_000
0.16
0 s
amte
ch_0
00
0.15
4 d
igita
lbar
riers
_001
0.12
1 d
erm
alog
_002
0.11
1 d
erm
alog
_003
0.10
5 v
igila
ntso
lutio
ns_0
01
0.07
5 c
yber
extr
uder
_001
0.07
0 r
anko
ne_0
00
0.06
2 i
d3_0
01
0.05
6 i
d3_0
02
0.05
2 n
euro
tech
nolo
gy_0
00
0.05
0 i
tmo_
002
0.04
8 r
anko
ne_0
02
0.04
3 n
euro
tech
nolo
gy_0
01
0.03
8 t
ongy
itran
s_00
1
0.03
5 v
ocor
d_00
1
0.03
4 i
nnov
atric
s_00
0
0.03
3 i
nnov
atric
s_00
1
0.03
0 v
isio
nlab
s_00
1
0.03
0 t
ongy
itran
s_00
2
0.02
9 3
divi
_000
0.02
7 n
tech
lab_
000
0.02
6 m
orph
o_00
0
0.02
5 n
tech
lab_
001
0.02
1 y
itu_0
00
0.01
3 v
ocor
d_00
2
Figu
re4:
For
the
visa
imag
es,d
etec
tion
erro
rtr
adeo
ff(D
ET
)cha
ract
eris
tics
show
ing
fals
eno
n-m
atch
rate
vs.f
alse
mat
chra
tepl
otte
dpa
ram
etri
cally
onth
resh
old,
T.Th
esc
ales
are
loga
rith
mic
inor
der
tosh
owm
any
deca
des
ofF
MR
.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 20is
it0
rank
0
rank
2
neur
0
ntec
0
ntec
1
ayon
0
vcog
1
vcog
2
vigi
0
vigi
1di
gi0
derm
2
derm
3
tong
1
tong
2
sam
t0
yitu
0
visi
1
voco
1
voco
2
3div
0m
orp0
inno
0
0.01
0.03
0.10
0.30 1
e07
1e
061e
05
1e
04Fa
lse
mat
ch r
ate
(FM
R)
False nonmatch rate (FNMR)
Dat
aset
: V
ISA
FN
MR
@ F
MR
=0.0
001
and
Alg
ori
thm
0.50
2 v
cog_
002
0.41
5 v
igila
ntso
lutio
ns_0
00
0.41
3 i
sity
ou_0
00
0.40
7 v
cog_
001
0.22
9 a
yoni
x_00
0
0.17
3 i
tmo_
001
0.16
0 d
igita
lbar
riers
_000
0.16
0 s
amte
ch_0
00
0.15
4 d
igita
lbar
riers
_001
0.12
1 d
erm
alog
_002
0.11
1 d
erm
alog
_003
0.10
5 v
igila
ntso
lutio
ns_0
01
0.07
5 c
yber
extr
uder
_001
0.07
0 r
anko
ne_0
00
0.06
2 i
d3_0
01
0.05
6 i
d3_0
02
0.05
2 n
euro
tech
nolo
gy_0
00
0.05
0 i
tmo_
002
0.04
8 r
anko
ne_0
02
0.04
3 n
euro
tech
nolo
gy_0
01
0.03
8 t
ongy
itran
s_00
1
0.03
5 v
ocor
d_00
1
0.03
4 i
nnov
atric
s_00
0
0.03
3 i
nnov
atric
s_00
1
0.03
0 v
isio
nlab
s_00
1
0.03
0 t
ongy
itran
s_00
2
0.02
9 3
divi
_000
0.02
7 n
tech
lab_
000
0.02
6 m
orph
o_00
0
0.02
5 n
tech
lab_
001
0.02
1 y
itu_0
00
0.01
3 v
ocor
d_00
2
Figu
re5:
For
the
visa
imag
es,
but
now
just
for
low
FM
R,
dete
ctio
ner
ror
trad
eoff
(DE
T)
char
acte
rist
ics
show
ing
fals
eno
n-m
atch
rate
vs.
fals
em
atch
rate
plot
ted
para
met
rica
llyon
thre
shol
d,T.
The
scal
esar
elo
gari
thm
icin
orde
rto
show
man
yde
cade
sof
FM
R.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 21
3div
0
digi
0
cybe
1
digi
1
rank
0
rank
2
neur
0
neur
1 nte
c0
ntec
1
mor
p0
visi
1
vigi
1
vigi
0
tong
1
tong
2
derm
2
derm
3
ayon
0
voco
1
voco
2
sam
t0
yitu
0
vcog
2
itmo1
itmo2
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
0.20
0.30
0.40 1
e05
1e
041e
03
1e
021e
01
1e+
00Fa
lse
mat
ch r
ate
(FM
R)
False nonmatch rate (FNMR)
Dat
aset
: M
UG
SH
OT
FN
MR
@ F
MR
=0.0
001
and
Alg
ori
thm
0.69
1 v
cog_
002
0.67
9 i
sity
ou_0
00
0.59
5 v
igila
ntso
lutio
ns_0
00
0.30
9 a
yoni
x_00
0
0.25
5 i
tmo_
001
0.24
0 d
erm
alog
_002
0.20
1 d
erm
alog
_003
0.18
3 d
igita
lbar
riers
_000
0.17
6 r
anko
ne_0
00
0.16
5 c
yber
extr
uder
_001
0.10
0 v
igila
ntso
lutio
ns_0
01
0.06
3 v
ocor
d_00
1
0.06
1 n
euro
tech
nolo
gy_0
00
0.04
7 i
tmo_
002
0.04
6 i
nnov
atric
s_00
0
0.04
5 n
euro
tech
nolo
gy_0
01
0.04
4 s
amte
ch_0
00
0.04
4 n
tech
lab_
000
0.04
3 i
nnov
atric
s_00
1
0.04
1 d
igita
lbar
riers
_001
0.04
0 t
ongy
itran
s_00
1
0.03
9 t
ongy
itran
s_00
2
0.03
7 3
divi
_000
0.03
6 i
d3_0
01
0.03
2 i
d3_0
02
0.03
1 r
anko
ne_0
02
0.03
0 n
tech
lab_
001
0.02
8 m
orph
o_00
0
0.02
4 v
isio
nlab
s_00
1
0.01
9 v
ocor
d_00
2
0.01
7 y
itu_0
00
Figu
re6:
For
the
mug
shot
imag
es,d
etec
tion
erro
rtr
adeo
ff(D
ET
)cha
ract
eris
tics
show
ing
fals
eno
n-m
atch
rate
vs.f
alse
mat
chra
tepl
otte
dpa
ram
etri
cally
onth
resh
old,
T.Th
esc
ales
are
loga
rith
mic
inor
der
tosh
owde
cade
sof
FM
R.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 22
3div
0
neur
0
rank
0
vcog
1
ayon
0
ntec
0
ntec
1
derm
2
sam
t0
tong
1
tong
2
vigi
0
voco
1
voco
2yi
tu0
mor
p0
inno
0
0.01
0.03
0.10
0.30 1
e05
1e
041e
03
1e
021e
01
Fals
e m
atch
rat
e (F
MR
)
False nonmatch rate (FNMR)
Dat
aset
: S
EL
FIE
SF
NM
R @
FM
R=0
.000
1an
d A
lgo
rith
m
1.00
0 i
sity
ou_0
00
0.66
6 v
cog_
002
0.64
3 v
igila
ntso
lutio
ns_0
00
0.42
7 v
cog_
001
0.36
0 a
yoni
x_00
0
0.19
9 i
tmo_
001
0.17
9 d
erm
alog
_002
0.17
0 d
igita
lbar
riers
_000
0.14
4 c
yber
extr
uder
_001
0.11
5 d
erm
alog
_003
0.11
2 d
igita
lbar
riers
_001
0.09
2 r
anko
ne_0
00
0.06
9 i
tmo_
002
0.06
9 v
ocor
d_00
1
0.06
3 s
amte
ch_0
00
0.06
3 t
ongy
itran
s_00
1
0.06
3 t
ongy
itran
s_00
2
0.05
8 v
igila
ntso
lutio
ns_0
01
0.05
5 3
divi
_000
0.05
2 n
euro
tech
nolo
gy_0
00
0.05
2 r
anko
ne_0
02
0.04
3 i
nnov
atric
s_00
1
0.04
0 i
d3_0
01
0.04
0 i
nnov
atric
s_00
0
0.03
7 i
d3_0
02
0.01
7 n
euro
tech
nolo
gy_0
01
0.01
4 n
tech
lab_
000
0.01
4 n
tech
lab_
001
0.01
4 v
isio
nlab
s_00
1
0.01
2 m
orph
o_00
0
0.01
2 v
ocor
d_00
2
0.01
2 y
itu_0
00
Figu
re7:
For
the
selfi
e-to
-por
trai
tco
mpa
riso
ns,d
etec
tion
erro
rtr
adeo
ff(D
ET
)ch
arac
teri
stic
ssh
owin
gfa
lse
non-
mat
chra
tevs
.fa
lse
mat
chra
tepl
otte
dpa
ram
etri
cally
onth
resh
old,
T.Th
esc
ales
are
loga
rith
mic
inor
der
tosh
owse
vera
ldec
ades
ofF
MR
.Cau
tion
:The
FNM
Rva
lues
here
are
opti
mis
tic
stat
emen
tsof
accu
racy
beca
use
the
imag
epa
irs
wer
eco
llect
edon
the
sam
eda
y.Th
isis
know
nac
ross
biom
etri
csto
give
bett
erac
cura
cy,a
ndis
oper
atio
nally
rele
vant
only
insp
ecia
lcas
es.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 23
vcog
1
vigi
0
3div
0
tong
1
sam
t0
rank
0
neur
0
ntec
0
ntec
1
derm
2
yitu
0
voco
1
mor
p0
digi
0
0.00
1
0.00
3
0.01
0
0.03
0
0.10
0
0.30
0 1e
051e
04
1e
031e
02
1e
01Fa
lse
mat
ch r
ate
(FM
R)
False nonmatch rate (FNMR)
Dat
aset
: W
EB
CA
MF
NM
R @
FM
R=0
.000
1an
d A
lgo
rith
m
0.68
9 i
sity
ou_0
00
0.55
7 v
cog_
002
0.40
0 v
igila
ntso
lutio
ns_0
00
0.30
1 v
cog_
001
0.17
1 a
yoni
x_00
0
0.10
9 d
erm
alog
_002
0.06
9 i
tmo_
001
0.04
5 d
igita
lbar
riers
_000
0.04
1 d
erm
alog
_003
0.03
6 v
ocor
d_00
1
0.03
6 i
tmo_
002
0.02
9 d
igita
lbar
riers
_001
0.02
9 c
yber
extr
uder
_001
0.02
1 s
amte
ch_0
00
0.02
0 r
anko
ne_0
00
0.01
6 v
igila
ntso
lutio
ns_0
01
0.01
0 t
ongy
itran
s_00
2
0.00
9 t
ongy
itran
s_00
1
0.00
7 m
orph
o_00
0
0.00
5 n
euro
tech
nolo
gy_0
00
0.00
3 i
d3_0
02
0.00
3 n
tech
lab_
001
0.00
3 n
tech
lab_
000
0.00
2 3
divi
_000
0.00
2 i
d3_0
01
0.00
1 i
nnov
atric
s_00
0
0.00
1 i
nnov
atric
s_00
1
0.00
1 n
euro
tech
nolo
gy_0
01
0.00
1 v
isio
nlab
s_00
1
0.00
0 r
anko
ne_0
02
0.00
0 y
itu_0
00
Figu
re8:
For
the
web
cam
-to-
port
rait
com
pari
sons
,det
ecti
oner
ror
trad
eoff
(DE
T)c
hara
cter
isti
cssh
owin
gfa
lse
non-
mat
chra
tevs
.fal
sem
atch
rate
plot
ted
para
met
rica
llyon
thre
shol
d,T.
The
scal
esar
elo
gari
thm
icin
orde
rto
show
seve
rald
ecad
esof
FM
R.C
auti
on:T
heFN
MR
valu
eshe
rear
eop
tim
isti
cst
atem
ents
ofac
cura
cybe
caus
eth
eim
age
pair
sw
ere
colle
cted
onth
esa
me
day.
This
iskn
own
acro
ssbi
omet
rics
togi
vebe
tter
accu
racy
,and
isop
erat
iona
llyre
leva
nton
lyin
spec
ialc
ases
.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 24
id3_
1
id3_
2
rank
0
rank
2
sam
t0
neur
0
visi
1yi
tu0
vigi
0
cybe
1
ayon
0di
gi0
digi
1
tong
1to
ng2
derm
2
derm
33d
iv0
ntec
0
ntec
1
voco
1
voco
2
mor
p0
inno
0
inno
1
vcog
2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9 1e
05
1e
041e
03
1e
021e
01
Fals
e m
atch
rat
e (F
MR
)
False nonmatch rate (FNMR)
Dat
aset
: W
ILD
FN
MR
@ F
MR
=0.0
01an
d A
lgo
rith
m
1.00
is
ityou
_000
0.93
itm
o_00
1
0.89
itm
o_00
2
0.88
vi
gila
ntso
lutio
ns_0
00
0.87
sa
mte
ch_0
00
0.86
vo
cord
_002
0.83
cy
bere
xtru
der_
001
0.80
ne
urot
echn
olog
y_00
0
0.78
m
orph
o_00
0
0.76
ay
onix
_000
0.71
di
gita
lbar
riers
_000
0.70
to
ngyi
tran
s_00
2
0.69
to
ngyi
tran
s_00
1
0.67
ne
urot
echn
olog
y_00
1
0.66
ra
nkon
e_00
0
0.66
di
gita
lbar
riers
_001
0.65
de
rmal
og_0
02
0.62
id
3_00
2
0.62
ra
nkon
e_00
2
0.62
vo
cord
_001
0.61
vi
gila
ntso
lutio
ns_0
01
0.59
de
rmal
og_0
03
0.59
vc
og_0
02
0.58
id
3_00
1
0.57
in
nova
tric
s_00
0
0.56
in
nova
tric
s_00
1
0.47
3d
ivi_
000
0.40
vi
sion
labs
_001
0.36
yi
tu_0
00
0.27
nt
echl
ab_0
00
0.24
nt
echl
ab_0
01
Figu
re9:
For
the
wild
imag
eco
mpa
riso
ns,d
etec
tion
erro
rtr
adeo
ff(D
ET
)ch
arac
teri
stic
ssh
owin
gfa
lse
non-
mat
chra
tevs
.fa
lse
mat
chra
tepl
otte
dpa
ram
etri
cally
onth
resh
old,
T.Th
esc
ales
are
loga
rith
mic
inor
der
tosh
owse
vera
ldec
ades
ofF
MR
.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 25
3div
0
digi
0
rank
0
visi
1
neur
0
ntec
0
ntec
1
tong
1
sam
t0
vcog
1
vigi
0
vigi
1
ayon
0
voco
1
voco
2
yitu
0
mor
p0
0.3
0.4
0.5
0.6
0.7
0.8 0
.001
0.01
00.
100
Fals
e m
atch
rat
e (F
MR
)
False nonmatch rate (FNMR)
Dat
aset
: C
hild
Exp
loit
atio
nF
NM
R @
FM
R=0
.000
1an
d A
lgo
rith
m
1.00
is
ityou
_000
0.98
de
rmal
og_0
02
0.89
vi
gila
ntso
lutio
ns_0
00
0.85
m
orph
o_00
0
0.84
de
rmal
og_0
03
0.84
ne
urot
echn
olog
y_00
0
0.84
ay
onix
_000
0.79
ra
nkon
e_00
0
0.77
di
gita
lbar
riers
_000
0.76
sa
mte
ch_0
00
0.76
vo
cord
_002
0.75
vc
og_0
02
0.75
to
ngyi
tran
s_00
2
0.74
to
ngyi
tran
s_00
1
0.73
vi
gila
ntso
lutio
ns_0
01
0.69
vo
cord
_001
0.69
vc
og_0
01
0.59
yi
tu_0
00
0.56
vi
sion
labs
_001
0.55
3d
ivi_
000
0.53
nt
echl
ab_0
00
0.47
nt
echl
ab_0
01
Figu
re10
:Fo
rch
ildex
ploi
tati
onim
ages
,de
tect
ion
erro
rtr
adeo
ff(D
ET
)ch
arac
teri
stic
ssh
owin
gfa
lse
non-
mat
chra
tevs
.fa
lse
mat
chra
tepl
otte
dpa
ram
etri
cally
onth
resh
old,
T.Th
esc
ales
are
loga
rith
mic
inor
der
tosh
owm
any
deca
des
ofF
MR
.Acc
urac
yis
poor
beca
use
man
yim
ages
have
adve
rse
qual
ity
char
acte
rist
ics,
and
beca
use
dete
ctio
nan
den
rollm
entf
ails
.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 26
rank
0
rank
1
digi
0
neur
0
yitu
0
ntec
0
ntec
1
tong
1
sam
t0
3div
0
vigi
0
vigi
1
vcog
1
vcog
2
derm
1
voco
1
voco
2
mor
p0
isit0
itmo1
ayon
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
13
1030
100
Ran
k at
whi
ch h
ighe
st s
corin
g m
ate
is fo
und
True positive identification rate (TPIR) aka "hit rate"
Dat
aset
: C
hild
Exp
loit
atio
nC
MC
@ R
ank=
1an
d A
lgo
rith
m
0.59
nt
echl
ab_0
01
0.52
nt
echl
ab_0
00
0.48
3d
ivi_
000
0.47
vi
sion
labs
_001
0.46
ne
urot
echn
olog
y_00
0
0.46
yi
tu_0
00
0.43
vc
og_0
01
0.39
vo
cord
_002
0.38
to
ngyi
tran
s_00
2
0.37
vo
cord
_001
0.36
vi
gila
ntso
lutio
ns_0
01
0.35
m
orph
o_00
0
0.34
ra
nkon
e_00
0
0.34
to
ngyi
tran
s_00
1
0.33
de
rmal
og_0
03
0.32
ra
nkon
e_00
1
0.29
sa
mte
ch_0
00
0.27
itm
o_00
1
0.27
di
gita
lbar
riers
_000
0.24
vc
og_0
02
0.23
ay
onix
_000
0.19
vi
gila
ntso
lutio
ns_0
00
0.18
is
ityou
_000
0.05
de
rmal
og_0
01
0.05
de
rmal
og_0
02
Figu
re11
:For
child
expl
oita
tion
imag
es,c
umul
ativ
em
atch
char
acte
rist
ics
(CM
C)s
how
ing
true
posi
tive
iden
tific
atio
nra
tevs
.ran
k.Th
isis
sim
ulat
ion
ofa
one-
to-m
any
sear
chex
peri
men
t-
see
disc
ussi
onin
sect
ion
4.2.
The
scal
esar
elo
gari
thm
icin
orde
rto
show
the
effe
ctof
long
cand
idat
elis
ts.
Acc
urac
yis
poor
but
muc
him
prov
edre
lati
veto
the
1:1
DET
sof
Fig.
10be
caus
ea
sear
chca
nsu
ccee
dif
any
ofa
subj
ects
seve
rale
nrol
led
imag
esm
atch
esth
ese
arch
imag
ew
ith
ahi
ghsc
ore.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 27
3div
i_00
0ay
onix
_000
cybe
rext
rude
r_00
1de
rmal
og_0
02de
rmal
og_0
03di
gita
lbar
riers
_000
digi
talb
arrie
rs_0
01id
3_00
1id
3_00
2in
nova
tric
s_00
0in
nova
tric
s_00
1is
ityou
_000
itmo_
001
itmo_
002
mor
pho_
000
neur
otec
hnol
ogy_
000
neur
otec
hnol
ogy_
001
ntec
hlab
_000
ntec
hlab
_001
rank
one_
000
rank
one_
002
sam
tech
_000
tong
yitr
ans_
001
tong
yitr
ans_
002
vcog
_002
vigi
lant
solu
tions
_000
vigi
lant
solu
tions
_001
visi
onla
bs_0
01vo
cord
_001
voco
rd_0
02
yitu
_000
0.01
0.03
0.05
0.10
0.20
0.30
0.01
0.03
0.05
0.10
0.20
0.30
0.01
0.03
0.05
0.10
0.20
0.30
0.01
0.03
0.05
0.10
0.20
0.30
0.01
0.03
0.05
0.10
0.20
0.30
0.01
0.03
0.05
0.10
0.20
0.30 1
e05
1e
041e
03
1e
021e
01
Fals
e m
atch
rat
e (F
MR
)
False nonmatch rate (FNMR)
Sex
F M
Rac
e
B W
Figu
re12
:Fo
rth
em
ugsh
otim
ages
,err
ortr
adeo
ffch
arac
teri
stic
sfo
rw
hite
fem
ales
,bla
ckfe
mal
es,b
lack
mal
esan
dw
hite
mal
es.
The
grey
lines
corr
espo
ndto
fixed
thre
shol
ds,s
how
ing
how
both
FNM
Ran
dFM
Rva
ryat
one
oper
atin
gth
resh
old.
Impo
rtan
t:M
any
ofth
epl
ots
will
naiv
ely
bere
adas
sayi
ngw
hite
sgi
ves
low
erer
ror
rate
sth
anbl
acks
beca
use
the
blue
trac
eslie
bene
ath
the
red
ones
.H
owev
er,t
his
ism
isle
adin
gan
din
com
plet
e:Th
egr
eylin
essh
owth
etr
aces
are
gene
rally
shif
ted
hori
zont
ally
.Th
usfo
rth
ede
rmal
og-0
01al
gori
thm
FNM
Rfo
rw
hite
sis
high
erth
anfo
rbl
acks
ata
fixed
thre
shol
dbu
t,at
the
sam
eti
me,
FMR
ishi
gher
for
blac
ks-
see
Figu
re17
.As
acce
ssco
ntro
lsys
tem
sal
mos
talw
ays
oper
ate
ata
fixed
thre
shol
d,th
ena
ive
inte
rpre
tati
onis
inco
rrec
t.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 28
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
M
M
MM
MM
M
M
MM
MM
MM
MM
MM
MM
MM
MM
M
M
MM
MM
M
M
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
M
M
MM
MM
M
M
MM
MM
MM
MM
MM
MM
MM
MM
M
M
MM
MM
M
M
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
MM
M
M
MM
MM
M
M
MM
MM
MM
MM
MM
MM
MM
MM
MM
3div
i_00
0ay
onix
_000
cybe
rext
rude
r_00
1de
rmal
og_0
02de
rmal
og_0
03di
gita
lbar
riers
_000
digi
talb
arrie
rs_0
01id
3_00
1id
3_00
2in
nova
tric
s_00
0in
nova
tric
s_00
1is
ityou
_000
itmo_
001
itmo_
002
mor
pho_
000
neur
otec
hnol
ogy_
000
neur
otec
hnol
ogy_
001
ntec
hlab
_000
ntec
hlab
_001
rank
one_
000
rank
one_
002
sam
tech
_000
tong
yitr
ans_
001
tong
yitr
ans_
002
vcog
_001
vcog
_002
vigi
lant
solu
tions
_000
vigi
lant
solu
tions
_001
visi
onla
bs_0
01vo
cord
_001
voco
rd_0
02yi
tu_0
00
0.01
0.02
0.04
0.06
0.08
0.10
0.20
0.30
0.40
0.60
0.80
0.01
0.02
0.04
0.06
0.08
0.10
0.20
0.30
0.40
0.60
0.80
0.01
0.02
0.04
0.06
0.08
0.10
0.20
0.30
0.40
0.60
0.80
0.01
0.02
0.04
0.06
0.08
0.10
0.20
0.30
0.40
0.60
0.80
0.01
0.02
0.04
0.06
0.08
0.10
0.20
0.30
0.40
0.60
0.80
0.01
0.02
0.04
0.06
0.08
0.10
0.20
0.30
0.40
0.60
0.80
1e
041e
03
1e
021e
01
1e
041e
03
1e
021e
01
Fals
e m
atch
rat
e (F
MR
)
False nonmatch rate (FNMR)
fmr_
no
min
al
0.00
001
0.00
003
0.00
010
0.00
030
0.00
100
0.00
300
Figu
re13
:Fo
rth
evi
saim
ages
,FN
MR
and
FMR
atsi
xop
erat
ing
poin
tsal
ong
the
DET
char
acte
rist
ic.
At
each
poin
ta
line
isdr
awn
betw
een
(FM
R,F
NM
R) M
AL
Ean
d(F
MR
,FN
MR
) FE
MA
LE
show
ing
how
whi
chse
xha
slo
wer
FMR
and/
orFN
MR
.The
M
labe
lden
otes
mal
e,th
eot
her
end
ofth
elin
eco
rres
pond
sto
fem
ale.
The
six
oper
atin
gth
resh
olds
are
sele
cted
togi
veth
eno
min
alfa
lse
mat
chra
tes
give
nin
the
lege
nd,a
ndar
eco
mpu
ted
over
alli
mpo
stor
pair
sre
gard
less
ofag
e,se
x,an
dpl
ace
ofbi
rth.
The
plot
ted
FMR
valu
esar
ebr
oadl
yan
orde
rof
mag
nitu
dela
rger
than
the
nom
inal
rate
sbe
caus
eFM
Ris
com
pute
dov
erde
mog
raph
ical
ly-m
atch
edim
post
orpa
irs
i.ein
divi
dual
sof
the
sam
ese
x,fr
omth
esa
me
geog
raph
icre
gion
(see
sect
ion
4.6.
1),a
ndth
esa
me
age
grou
p(s
eese
ctio
n4.
6.2)
.
2017/08/25 13:00:41 FNMR(T) False non-match rateFMR(T) False match rate
-
FRVT - FACE RECOGNITION VENDOR TEST - VERIFICATION 29
3div
i_00
0ay
onix
_000
cybe
rext
rude
r_00
1de
rmal
og_0
02de
rmal
og_0
03di
gita
lbar
riers
_000
digi
talb
arrie
rs_0
01id
3_00
1id
3_00
2in
nova
tric
s_00
0in
nova
tric
s_00
1is
ityou
_000
itmo_
001
itmo_
002
mor
pho_
000
neur
otec
hnol
ogy_
000
neur
otec
hnol
ogy_
001
ntec
hlab
_000
ntec
hlab
_001
rank
one_
000
rank
one_
002
sam
tech
_000
tong
yitr
ans_
001
tong
yitr
ans_
002
vcog
_002
vigi
lant
solu
tions
_000
vigi
lant
solu
tions
_001
visi
onla
bs_0
01vo
cord
_001
voco
rd_0
02
yitu
_000
0.1
0.3
0.5
0.1
0.3
0.5
0.1
0.3
0.5
0.1
0.3
0.5
0.1
0.3
0.5
0