Face Masks Increase Face Recognition Errors Says NIST
COVID-19 has led to widespread facemask use, which as IPVM testing has shown drastically reduces face recognition accuracy.
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 *******
***** **** ********** ****** ***** *** over **% ******** **** **** ***** worn, **** ********** *** ***** ******** fall **** ~**% ******* ***** ** ~**% **** masks.
*******, ******** ********** *** ******* ******** the **** ***** **** ***** **** having * ***** ****** ** *** true ********* ****, ** ***-***** *** want ** ****** ********** ** ******** for *****. *******, ** ** ** safe, ** *** **** ***** *** users ** ***** ***** *****, ******** their ****, ***** ***** *******, ** exposing *** **** *** ******** *** negative ****** ** * **** **** by **% ** ****.
**** **** *** ******* ********** ********** algorithms *** **** ** ***** *** research ******** **** *** *** ******* what *** ******* ** ******** ******* in **********.
******:**** *** ****** ***** *******************, **** ** ***** **** ********* verify **+% *** ******* **+% ** people ******* **** *****.
*******
*** **** ***** ******* ** ***** Non-Match **** *** ***** ***** ****, which ** ********** *** ***** *********** that *** *** ****** *** ***-*****. They ****** ** *** ****** ** the ********* ** *****, ** ** algorithm ***** **** **.*% ******** ** 99% **** **** ***** ***** ** 10 ***** ***** ** ***** ******, but ** ********* ***** **** **% accurate ** **% ***** **** ** 5 ***** *****. **** ** **********, so ** *** ********* ** ***** of *** **** ************ **** (*** percent ** ********* ********* ******** *** images ** *** **** ****** ** a *****), *** *** **** ********* rate (*** ******* ** ********* ********* labels *** ********* ****** ** *********).
Study ********
**** **** ~* ******* ****** ******** images ***** **** ******* ** ******* with ~* ******* **** *********** ****** for ************/ *:* ******** (*.*. ** the ****** ******** ***** ******* *** visa *****). *** **** *********** ****** were ****-******* *** ***** ** *********** offices **** * ***** **********. **** are ******* ** ***/*** ***** *.*. JPEG *** *** *** * *** pixels. *** ****** ******** ****** *** not ****** **** *** ***/*** *****-* full-frontal ***** ********* *** ** **** cases **** ****** *********.
**** **** ***** ******* “*****” ** the ****** ******** ******. **** ****** 2 ****** (*****-**** *** *****), * variations (**** *** *****), *** * levels ** **** ******** (***/**** ************, medium/nose ********* *******, ****/**** ****** ** to ***-*****, ********** ******** ****).
**** **** ******** *** ****** ******** images (**** ********* ******* **** ************ applied) **** *** **** *********** ******, testing **** ** *** ** **********. NIST **** * ***** *********, **** led ** * ** ***,*** ******** images ***** ********, ** ****** ** something *** * ***** ** ***.
****, **** *** *** **** **** combination ** ***** ********* *** ** time ***********, *** ******* **** ********** were ****** ** ****** * **** when ***** **** ****.
**** *** ********* ******* ***** ************** *** **** *******.
Results **** * **** **** **** *** ****** **** ********
******: ***** ******* *** ***** ** pre-COVID **********.
***** ********* *** **** ************ ****, as **** ** ********* *** **** an ********* ****** ** ****** * face, ****** **** ********** ******** ********** error-prone. ******** ******* ** *** *****, but *** ****** ** ****** ********. While *********'* *** **********'* **** ************ rate **** ********* ** *.* *** 2 ********** ****** ************, *******'* **** verification **** ********* ** **.* ********** points, ** *** ***** ***** **********:
*** ***** ***** ***** *** **** data ** ******* **** *** * selected ****** ** ********** ****** ** NIST:
**** ***** ********** ********* ******* ******** ** *********'* ***-********** **** * **** ****. *******, no ********* *** **** ****** **** it. ****, ** ********* **** * masked ******* ** ***** ** *** best ********** **** * ******** ***.
Effects ** ********* **** ******
** *** *****, ***** ********* ********. Wider ***** *** * ****** ****** than ***** **** (*** *****).****, ***** *******, **** ***** ***** having * ******** ****** ****** **** blue (*******, **** *** *** **********).
*** ******* ****** *** **** ********, which ** **** *****, *********** ** over **** *** *** ******** ** true ************ ********.
Implications *** ***-*****
***-***** **** ** ********** ** ***** face *********** ********* *** ***** ****** with ********** **** **** ***. **** of *** **** ********** ********** ** unmasked ***** ****** ********* ********** ** masked ********, ***** ***** ********** ******** high-levels ** ********, ** ***-***** **** understand *** ****-*** ******* ***** ********** algorithm.
Adjusting ********** ** ***** ******
******** ********** *** **** ***** *** effect ** ****-***** ** ********** ******* ************ ******** **** ******* ******* ** ***** acceptances, ** ***** ** *** ******* below.
Exposing *****
** ** ** ****, ** *** make ***** *** ***** ** ***** their ***** ******** ***** ****, ** this *** ******* ******** *** **** verification ****.
***********
**** **** *** ******* ********** ********** algorithms *** **** ** ***** *** research ******** **** *** *** ******* what *** ******* ** ******** ******* in **********.***** *** ******* *** ***** ***********, it ***** **** ***-***** ****** *********** use *** ****** ** *** *** a ***** ******** **** ** ******** by ****-***.
**** ** *** ********** **** ********* on ****-******* ******** ** ******** *** mask-wearing ********. ********** *** ************* ******* the *********** ** ** ********* ** masked ********, *** ********* *** **** released ******** ******* ** ******* ****.
****, *** ***** **** *******. ***** this ****** *** ********** *********** *** was ********* *** *** ****** ** testing, ** ***** *** ******* *** not ******* ********* *******. ***** *** many ***** ** ****-***** **** **** not ******, **** ** ***** *** have ********* *******. ****, *** ***** or ******* ** * **** *** impact *** ***** ***** ** *** subject, ** ********** ***** ** ******** shadows.
*******, **** *** *** **** *** effect ** ********* ******* * ****** image.******* ****** **** ***** ** *** visa *********** ****** *****’** ********* *** negative ******* ** ******* * **** mask.
** **** ******* ** ******** ****'* future **** *********** ****** **** ******* on *** ******* ** ****-*****.
***** *** **** *** **** ***** are ********* *** *** **** **** I *** ** *** ** **** report ***** **** ******** **** ***** the ********* ** *** *********** *** make. ******* *****? ** ***** ********** in *** **** ******? ***** ** well ** ******* ***** ******* ** a **** *******. *****! * **** my *** ******* ****.
** ** *****, *** *********** ****** is *** *** **** ** ******* (fully ** *********). ** * ****** wear * **** ******** **** .... sorry ******* *** *** ** ********.
** ****** **** *** **** **** ON *** ********** ** ***** ****.
***** *** * ****** ** ****** users *** ***** *********** ****** **** keep ** **** ****. *****, ** how ** *** *********** ******* (*** a ****** *****) **** **** ******* does *** **** * **** **? Are *** ******** ***** *** **** as ****** ** **** **** ******* (or *** ***** ******* ** ***** size *********) *** ** *** ************ of *** ******** ****? ******, ** how **** *** ************ ** * mask ********* **** ***** (****, ******)? It *** ***** ** * *** of **** *** * *** ** real *** ********* ******** **** ** get *** ***** *** ** ********** levels. ****, *** ********** ** *** threshold ** * ***** ********, ********** if ******* ** ***** ****-***** ********* where *** ************ ** * ***** positive ******* *****-****** ******* ** ********** (as ******* ** ** ******** ********).
** **** ** *********** ** *** how ********** ********* ** **** **** masks ** **** ******* ** ********** submitted ****** ***** *** ****** ** unmasked *****. **** **** ****** **** worse *******, *** **** ********** **** pretty ********* ******* **********, ** ***** companies **** ** **** ** ******* high ******** **** ***** ** ** well.
*** *'* *** **** **** ********** bias *****, *******'* ******** ****** ** ********* ********** ******, ***********-*****, *** ******-***** bias ******* ********** ******* *****.
****, "** ******** ********" ***** ** ********** problematic ******* ***** ********* ************* **** watchlist ****.
**'* **** *** **** ** ** at **** *** *** ********** ***** is **** ** *.*% ** ** on *** *** ******* (****** ***** ***** ********* ******** ** Bias *********) *** ******* ***** **** *******.
*** * ***** **** *** ** the ******** ********. ****'* ***** *** really *** ******** *** ***** *********.
******* ** **** - *** ******** - *** ***** *** *** ****** of ******** ********* (***** ****** ****) has *** ******** ********* ****** **** 1/100,000 (*.*****) ** */*** (.**). ** don't ********* *** * ***** ****** in *** ********* ****** ******* ****** on *** ***. *** **** * typo ** *** *****? * *** a ****** *** ****** ** *** NIST ****** *** ****'** ****** *** source **** ** *****.
** ***'* ********* *** * ***** change ** *** ********* ****** ******* impact ** *** ***.
*** ******* ** * ****** ***** there, *** **** *****'* ******* *** actual **********, **** * ***** ** FNMR ** ***, ** ***** *********** to *** ****** ** *** *** threshold. ***, * ***** ****** ** FMR ****** **** * ***, ***** is *** ** ***** ** ********* it ** ** ******** ** *** error ** * ********** ******, ***** I ***** ** * **** *********** metric.
***, *** ***** ** **** ** from ** ** **** **. **'* a ****** **** ** ***, *** the **** *** *** *** ***** for ***** ********* ** ********* ********** (not *******).
*'** ******* *** ******* *********** *** ****** ***** *******************, **** ** ***** **** ********* verify **+% *** ******* **+% ** people ******* **** *****.
***** *** ******* ************* ******* **** the *****. * ****** *** **** would ****.