Herta Facial Recognition Plus Masks Tested

By Rob Kilpatrick, Published Aug 19, 2020, 09:54am EDT

Masks increase face recognition errors, but facial recognition developer Herta is claiming their algorithm gives priority to the upper part of the face, offering accurate recognition even with masks on.

IPVM Image

We tested Herta's BioSurveillance facial recognition for over a month. In this report, we examine:

  • How do masks impact facial recognition confidence?
  • Do masks increase false positives and negatives?
  • How do masks impact recognition at harsh angles or in poor lighting?
  • How does recognition without masks compare to wearing them?
  • How simple is Herta's UI to operate?

Executive *******

** *** *****, ***** significantly ******* *****'* ******** confidence (**** **-** ** 20-40) *** *** *** result ** *** ***** positives ****** *** ****** of * ***** ** testing **** * ********* of ****** *** ********.

*******, ******** ********** ********* requires ***** *** *****, as ******* ** *** high (~**+) ******** ** missed ********** ** ***** subjects, ***** ******* ** too *** (~** *** lower) ******** ** ***** positives.

************, **** ***** **, increasing ******** (>**°) ** angle ** ********* ** reducing ***** ****** (****** color *** ***** ** with ** **) ******** in **** ****** *********** rates *** ** *** recommended.

******* *****, *********** *** solid **** ******** ********** recognized ** **** ***** even ** ***** ****** (up ** **° **********) and ** *** ***** and ** ** ******** angles (Herta *** *** ****** from *********** ******* ** false ********** ** ********* objects **** ** ***** facial *********** ********* ** have ******, **** *********************.

Relatively ***** ****** ****

** *** *****, ***** observed ******* ******* ****** subjects, ********* **** *********, employees *** ******** ** nearby **********, ** **** as ******* ***** ** surveillance ******* ***** **** various ******* ** *** internet.

*******, ***** ********** *** relatively ***** ******** ** many ************ **** *** is ****** ** ** implemented, **** ** ******* with ******* ******* ********, casinos **** ********* ** visitors, *** **** ******* with **** ** ********* of ****** ********. ******* in ****** *********** *** vary, **** ***** *********** issues **********.

Masks ************* ****** **********

**** * ****** ******* a **** ** ********, confidence ***** ************* **** ~80-90 ** ~**-**.

IPVM Image

Accurate *********** **** *****

******* **** ******* ********* in **********, *** ******** did *** ******* *** false ********* (*** ******* mistaken *** *******) ********** a ***** ** *******, with * ********* ** over *** ******** ** varying ************.

IPVM Image

Missed ********** **** *****

*******, ** ***** **** Herta ****** ** ********* known ******** ******* ***** when ***** ***** *********** threshold ** **.

IPVM Image

** ***** **** ******** the ********* ** ** increased *** ****** ** recognitions ** ****** ******* masks ***** *** ****** any ***** ************. ***** below **** ********* ********** false *********, **** ******** mistakenly ********** ** ***** on *** *********.

IPVM Image

Standard / ***-**** ***********

****** ~* ***** ** testing ***** **** ** false ************ ***** ****** suggested ********** ** **, this ******** * ******* of ****** *** ****** in * ********** ******, warehouse, *** ******* ******* lot **** ******** ** subjects ******* *******.

IPVM Image

Subjects ********** **** **** ********** ** **** *****

** * **** ***** setting ******** **** ********** with * ********** ** ~80 - **, **** confidence ******* ** ~** with ****** ******** ******* of ~**' *** ********* of ~**(*******).

IPVM Image

******** **** ********** *********** at * ** (******) angle **** ~** **** confidence ** **** *****.

IPVM Image

Reduced ********** **** **

********** ******* ** ~** in * **** ***** (~0.02 ***) **** **, resulting ** ************ ** ~8' ******** ****** **** ~22° ** ******** ***** the *********** ** ********** threshold.

IPVM Image

******* ** **° ******** in ***** ********** **** the ** ********* *** missed ************ ******** **** at *** ******** ******* of ~*' *** ********* of ~**°.

IPVM Image

Reduced ********** ** *** ***** ***** (~* ***)

********** ** ************ *** reduced ** * *** light ***** ** ~** points ********* ** **** recognitions ** ***** ******** heights ** ~*' *** downtilts ** ~**° **** walking ****** *** ******.

IPVM Image

Confidence **** ** *** *****

********** ******* ** ~** when *********** ***** ** a *** *****. **** combined **** *****, **** regularly *** ** ****** recognitions.

IPVM Image

Sunglasses / **** ******

******** ******* ********** ** the ***** **** ********** at ~** ***** ********** drop **** * **** in **** *****.

IPVM Image

******* * *** ** the ***** ******** ** ~10 - ** ********** loss *** ***** ******* recognition ** ****** ******** heights ** ~**' *** downtilts ** ~**°.

IPVM Image

*******, * *********** ** sunglasses *** **** ******* confidence ** ~**, ********* in ****** ************ ***** Herta ******* **********.

IPVM Image

**********

***** ** *** ****** is ****** **** *********, adding * ****** ** the ****** *** ** done ** ********* * detection ** *** ******** face ** ********* ** image (******* ** * passport *****) ** *** subject ** *** ******.

** *** ***** **** to *** ****** ** the ****** ** **** a ******** ********, *********** needs ** ** ****.

Simple ** **********

****** ******* ** ***** that *** ** ** Herta *** ****** ** navigate. ****** ******* *** be **** ** *** main ******, ** **** as ******* ********** *** recognitions.

No ***** ********* / *** ********

**** * ****** ** recognized ***** ** ** video ** *** ***** available, **** * ***** image ** *** ****** and *****.

IPVM Image

***** ** **** **** to ** ****** ** a *** ***** *** such ** ********* ***** below, ****** *** ********** as **** ** *** video **** *********** **** the ***********.

IPVM Image

*******

***** *************** *** ** MSRP ** *,*** € / ~$*,*** *** *** channel, ***** ******** * database **** ** ** to *,*** ********. ********* database ***** *** ********* as ********** ********. **** MSRP ** ********** ** slightly ***** **** ***** facial *********** ********* ** have ******, **** ** Anyvision (~$*,*** *** ****). However, ***** ********** **** MSRP ******* ** **** a *********, ** **** contracts *** ****** *** project, ********* ********* ** additional ********.

**** ******* **** **** not ******* ****** ******** needed, ****** ******* $*,***+, depending ** ****. * dedicated *** ** ********, but ***** ************ ****** by ****** *****. *** four *******, * *** 1660 *** ** ****, an ********** $***-*** ***.

Versions ****

*** ********* ******** **** used ****** *******.

  • *****: **.*.**

Related: **** ********* *******

***** ** *** ***** full ****** *********** ****** we **** ****** ******** accurate ********* **** *****, though ** ****** **** developers ** **** ******* claims *** ** ********** mask *** ****** *** COVID-19 ********.

*******, ** **** **** tested**** ********* *********, ********* wearing *****, *** ******* ********* claiming **** **** *********, including:

Comments (3)

** *** ********* ** about *,*** € ********* or *** ****?

***, **** ****, *** should **** ***** ***** for *** *********** ***** of *** *********, *.*. what *** *** **** getting ** **** ********* value, ** *** ** vendor ***** **** **** based ** ***** ******** studies. ***** **** ******** was **** *** *******, the ***** ***** ** reported ***'* **** ****. Somewhat ********* ** *** scenario, *** ****** ******** would ****** *** ****** FAR ** ** ***** of *** *.**% (*.*. false ******** ** * out **** *****) *** access ******* *** ****** than **** ** **** up ******** ***** ******* at * ****** ** the ************ *** ****. Thus ** *** **** a ******** ** *** 10K ******, **** *** algorithm ********* ****** ** in *** ***** ** 0.01%/10000 = ***-* (****, slightly ****** **** ****, but ****'* * ****** estimate).

Agree
Disagree
Informative: 2
Unhelpful
Funny

***** ****** *** *********. There's * **% ******** maintenance ****** *** ****, optional. ******** ******** *** not ******** ******* *** maintenance ******.

********* ***, ** **** gotten ** **** ******** vendors, ********* *****. *******, we ***'* *** **** faith ** ** ** all. ******* ***** ********** are ********* ***** *** NIST *******, ***** **** much ****** ********** ****** than ********* ******** ** surveillance *******. *** *******, the **** ****-** ******* is ****** ******* ******, so *** ***** *** foot ****** *** *** uncommon, *** ******** *** forced ** **** ** the ******, ***** ** much ****** ** ******* in *** ****.

******* * ******* ***** from **** *******:

IPVM Image

** * *******, ****** high ******* ************ *****:

IPVM Image

*** ******* ** ***** resolution (**-** ****** ****** the **** ******* ** 200) *** ****** *** harsher.

******* **** ******** *'** seen ***'* ******* ***** realities, ***** ******* ********* using **** *** * rough **** **** **** a ****** ****** ****.

Agree: 1
Disagree
Informative: 2
Unhelpful
Funny

****** *** *** ****** Ethan! ** *** *** into **** ******* *** will *** **** *** FRR ** ********* * function ** *** ******** (i.e. *** ****** **** the *** ** *** will **** **** *** false *******, *** *** images **** *** *** at *** ****** **** have **** ***** *******). However, ** *** *** the ********* ** *** 20 **** (*** * good ********* ** *****:)) the ***** ****** **** will ** *** **** regardless ** ***** *******. NIST ****** *** **** in **** ** ***** studies (*** ********* ** the ********* **. *** is *********** ** *** NIST ****** ** ***** grey ***** ********** *** curves *** ********* *********, which ******* ****** ** vertical - **** *** perfect, *** **** **** at ***** **** ****** the **** ***** ** magnitude **** **** ******* tell ***).

Agree
Disagree
Informative: 1
Unhelpful
Funny
Login to read this IPVM report.
Why do I need to log in?
IPVM conducts unique testing and research funded by subscriber's payments enabling us to offer the most independent, accurate and in-depth information.
Loading Related Reports