Brendan Klare（ROC）：人脸识别的挑战与机遇.pdf

1、Presented by:Brendan F.Klare,Ph.D.03 April 2025Challenges and Opportunities in Face Recognition22024 ROC|All Rights ReservedSummary of Face Recognition Progress Last IFPC,we discussed the continued exponential reductions in error ratesAre these still occurring?Top 10 Algorithms newer than 3 years:To

2、p 10 Algorithms older than 3 years:Mugshot accuracy largely unchanged the last 3 yearsSource:https:/pages.nist.gov/frvt/html/frvt11.html(Accessed on 4-1-2025)Top 10 Algorithms newer than 3 years:Top 10 Algorithms older than 3 years:Visa Border:1.5x reduction in error rate in 3 yearsSource:https:/pag

3、es.nist.gov/frvt/html/frvt11.html(Accessed on 4-1-2025)Top 10 Algorithms newer than 3 years:Top 10 Algorithms older than 3 years:Border Kiosk:1.3x reduction in error rate in 3 yearsSource:https:/pages.nist.gov/frvt/html/frvt11.html(Accessed on 4-1-2025)Top 10 Algorithms newer than 3 years:Top 10 Alg

4、orithms older than 3 years:Yaw 45 degrees:Over 3x error rate reductionSource:https:/pages.nist.gov/frvt/html/frvt11.html(Accessed on 4-1-2025)72024 ROC|All Rights ReservedVendor Error Rate Reduction Over Time2017-2022:Exponential error rate reduction2023 to now:linear reduction in error rate82024 RO

5、C|All Rights ReservedIs progress is slowing down,or are certain datasets getting“solved”in constrained use-cases?What are these remaining errors in the mugshot dataset?Identical twins?Quality issues?Ground truth errors?Confounding dopplegangers?How consistently are these top 10 algorithms all gettin

6、g the wrong answer on the same comparisons?If all these extremely accurate algorithms are wrong on the same comparisons,are they wrong,or are us humans wrong?82024 ROC|All Rights ReservedChallengesTop 10 Lowest Error Rates on Mugshot 1:1Perhaps issue is 1:1 FR benchmarks need to measure FNMR at lowe