Face Masks Increase Face Recognition Errors Says NIST

By Zach Segal, Published Aug 04, 2020, 10:24am EDT

COVID-19 has led to widespread facemask use, which as IPVM testing has shown drastically reduces face recognition accuracy.

IPVM Image

Now NIST (National Institute of Standards and Technology) is analyzing the effect of masks on face recognition algorithms through their Face Recognition Vendor Tests (FRVT). We will explore the results of their recently released FRVT 6A, on face recognition algorithms submitted before COVID-19 (without the expectation of being used on masked subjects) to explain the impact face masks have on face recognition algorithms and answer what the study means for users.

Executive *******

***** **** ********** ****** still *** **** **% accuracy **** **** ***** worn, **** ********** *** their ******** **** **** ~**% ******* ***** ** ~50% **** *****.

*******, ******** ********** *** greatly ******** *** **** match **** ***** **** having * ***** ****** on *** **** ********* rate, ** ***-***** *** want ** ****** ********** to ******** *** *****. Finally, ** ** ** safe, ** *** **** sense *** ***** ** lower ***** *****, ******** their ****, ***** ***** scanned, ** ******** *** nose *** ******** *** negative ****** ** * face **** ** **% or ****.

**** **** *** ******* submitting ********** ********** *** many ** ***** *** research ******** **** *** not ******* **** *** company ** ******** ******* in **********.

******:**** *** ****** ***** submitted**********, **** ** ***** will ********* ****** **+% and ******* **+% ** people ******* **** *****.

*******

*** **** ***** ******* on ***** ***-***** **** and ***** ***** ****, which ** ********** *** gives *********** **** *** not ****** *** ***-*****. They ****** ** *** change ** *** ********* of *****, ** ** algorithm ***** **** **.*% accurate ** **% **** face ***** ***** ** 10 ***** ***** ** their ******, *** ** algorithm ***** **** **% accurate ** **% ***** only ** * ***** worse. **** ** **********, so ** *** ********* in ***** ** *** true ************ **** (*** percent ** ********* ********* verifies *** ****** ** the **** ****** ** a *****), *** *** true ********* **** (*** percent ** ********* ********* labels *** ********* ****** as *********).

Study ********

**** **** ~* ******* border ******** ****** ***** with ******* ** ******* with ~* ******* **** Application ****** *** ************/ 1:1 ******** (*.*. ** the ****** ******** ***** matched *** **** *****). The **** *********** ****** were ****-******* *** ***** in *********** ******* **** a ***** **********. **** are ******* ** ***/*** 10918 *.*. **** *** are *** * *** pixels. *** ****** ******** images *** *** ****** with *** ***/*** *****-* full-frontal ***** ********* *** in **** ***** **** highly *********.

**** **** ***** ******* “masks” ** *** ****** crossing ******. **** ****** 2 ****** (*****-**** *** black), * ********** (**** and *****), *** * levels ** **** ******** (low/nose ************, ******/**** ********* covered, ****/**** ****** ** to ***-*****, ********** ******** nose).

IPVM Image

**** **** ******** *** border ******** ****** (**** different ******* **** ************ applied) **** *** **** Application ******, ******* **** of *** ** **********. NIST **** * ***** threshold, **** *** ** 1 ** ***,*** ******** images ***** ********, ** decide ** ********* *** a ***** ** ***.

****, **** *** *** test **** *********** ** every ********* *** ** time ***********, *** ******* some ********** **** ****** to ****** * **** when ***** **** ****.

**** *** ********* ******* newer ************** *** **** *******.

Results **** * **** **** **** *** ****** **** ********

******: ***** ******* *** based ** ***-***** **********.

***** ********* *** **** verification ****, ** **** as ********* *** **** an ********* ****** ** detect * ****, ****** many ********** ******** ********** error-prone. ******** ******* ** all *****, *** *** change ** ****** ********. While *********'* *** **********'* true ************ **** **** decreased ** *.* *** 2 ********** ****** ************, Samsung's **** ************ **** decreased ** **.* ********** points, ** *** ***** below **********:

IPVM Image

*** ***** ***** ***** the **** **** ** tabular **** *** * selected ****** ** ********** tested ** ****:

IPVM Image

**** ***** ********** ********* ******* ******** ** Deepglint's ***-********** **** * **** worn. *******, ** ********* has **** ****** **** it. ****, ** ********* with * ****** ******* is ***** ** *** best ********** **** * maskless ***.

Effects ** ********* **** ******

** *** *****, ***** decreased ********. ***** ***** had * ****** ****** than ***** **** (*** style).****, ***** *******, **** black ***** ****** * slightly ****** ****** **** blue (*******, **** *** not **********).

*** ******* ****** *** nose ********, ***** ** many *****, *********** ** over **** *** *** decrease ** **** ************ accuracy.

Implications *** ***-*****

***-***** **** ** ********** if ***** **** *********** solutions *** ***** ****** with ********** **** **** use. **** ** *** best ********** ********** ** unmasked ***** ****** ********* inaccurate ** ****** ********, while ***** ********** ******** high-levels ** ********, ** end-users **** ********** *** mask-use ******* ***** ********** algorithm.

Adjusting ********** ** ***** ******

******** ********** *** **** limit *** ****** ** face-masks ** ********** ******* ************ ******** **** ******* ******* on ***** ***********, ** shown ** *** ******* below.

IPVM Image

Exposing *****

** ** ** ****, it *** **** ***** for ***** ** ***** their ***** ******** ***** nose, ** **** *** greatly ******** *** **** verification ****.

***********

**** **** *** ******* submitting ********** ********** *** many ** ***** *** research ******** **** *** not ******* **** *** company ** ******** ******* in **********.***** *** ******* *** still ***********, ** ***** that ***-***** ****** *********** use *** ****** ** see *** * ***** solution **** ** ******** by ****-***.

**** ** *** ********** were ********* ** ****-******* subjects ** ******** *** mask-wearing ********. ********** *** significantly ******* *** *********** of ** ********* ** masked ********, *** ********* may **** ******** ******** updates ** ******* ****.

****, *** ***** **** digital. ***** **** ****** for ********** *********** *** was ********* *** *** volume ** *******, ** means *** ******* *** not ******* ********* *******. There *** **** ***** of ****-***** **** **** not ******, **** ** which *** **** ********* effects. ****, *** ***** or ******* ** * mask *** ****** *** photo ***** ** *** subject, ** ********** ***** or ******** *******.

*******, **** *** *** test *** ****** ** verifying ******* * ****** image.******* ****** **** ***** to *** **** *********** images *****’** ********* *** negative ******* ** ******* a **** ****.

** **** ******* ** covering ****'* ****** **** Recognition ****** **** ******* on *** ******* ** face-masks.

Comments (9)

***** *** ******* ************* missing **** *** *****. I ****** *** **** would ****.

** ** *****, *** significant ****** ** *** the **** ** ******* (fully ** *********). ** a ****** **** * mask ******** **** .... sorry ******* *** *** do ********.

** ****** **** *** face **** ** *** percentage ** ***** ****.

***** *** * ****** of ****** ***** *** their *********** ****** **** keep ** **** ****. First, ** *** ** the *********** ******* (*** a ****** *****) **** when ******* **** *** have * **** **? Are *** ******** ***** the **** ** ****** or **** **** ******* (or *** ***** ******* or ***** **** *********) due ** *** ************ of *** ******** ****? Second, ** *** **** the ************ ** * mask ********* **** ***** (race, ******)? ** *** taken ** * *** of **** *** * mix ** **** *** synthetic ******** **** ** get *** ***** *** to ********** ******. ****, the ********** ** *** threshold ** * ***** argument, ********** ** ******* is ***** ****-***** ********* where *** ************ ** a ***** ******** ******* stack-ranked ******* ** ********** (as ******* ** ** alarming ********).

** **** ** *********** to *** *** ********** submitted ** **** **** masks ** **** ******* to ********** ********* ****** COVID *** ****** ** unmasked *****. **** **** should **** ***** *******, but **** ********** **** pretty ********* ******* **********, so ***** ********* **** be **** ** ******* high ******** **** ***** on ** ****.

*** *'* *** **** what ********** **** *****, but****'* ******** ****** ** bias***** ********** ******, ***********-*****, and ******-***** **** ******* algorithms ******* *****.

****, "** ******** ********" ***** be ********** *********** ******* error ********* ************* **** watchlist ****.

**'* **** *** **** to ** ** **** but *** ********** ***** is **** ** *.*% or ** ** *** RFW ******* (****** ***** ***** ********* Industry ** **** *********) *** ******* ***** each *******.

*** * ***** **** you ** *** ******** scenario. ****'* ***** *** really *** ******** *** false *********.

******* ** **** - one ******** - *** table *** *** ****** of ******** ********* (***** Sertis ****) *** *** imposter ********* ****** **** 1/100,000 (*.*****) ** */*** (.03). ** ***'* ********* see * ***** ****** in *** ********* ****** minimum ****** ** *** FAR. *** **** * typo ** *** *****? I *** * ****** for ****** ** *** NIST ****** *** ****'** missed *** ****** **** in *****.

** ***'* ********* *** a ***** ****** ** the ********* ****** ******* impact ** *** ***.

*** ******* ** * little ***** *****, *** NIST *****'* ******* *** actual **********, **** * graph ** **** ** FMR, ** ***** *********** to *** ****** ** FMR *** *********. ***, a ***** ****** ** FMR ****** **** * lot, ***** ** *** we ***** ** ********* it ** ** ******** in *** ***** ** 3 ********** ******, ***** I ***** ** * more *********** ******.

***, *** ***** ** took ** **** ** on **** **. **'* a ****** **** ** see, *** *** **** and *** *** ***** for ***** ********* ** different ********** (*** *******).

****. ****** *** *** context.

*******. *'* **** *** enjoyed *** *******.

*'** ******* *** ******* because**** *** ****** ***** submitted**********, **** ** ***** will ********* ****** **+% and ******* **+% ** people ******* **** *****.

Read this IPVM report for free.

This article is part of IPVM's 6,736 reports, 909 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports