Herta Facial Recognition Plus Masks Tested

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

***** ******** **** *********** ******, *** ****** *********** **************** ******** ***** ********* ***** ******** to *** ***** **** ** *** face, ******** ******** *********** **** **** masks **.

IPVM Image

** ****** *****'* *************** ****** *********** for **** * *****. ** **** report, ** *******:

  • *** ** ***** ****** ****** *********** confidence?
  • ** ***** ******** ***** ********* *** negatives?
  • *** ** ***** ****** *********** ** harsh ****** ** ** **** ********?
  • *** **** *********** ******* ***** ******* to ******* ****?
  • *** ****** ** *****'* ** ** operate?

Executive *******

** *** *****, ***** ************* ******* Herta's ******** ********** (**** **-** ** 20-40) *** *** *** ****** ** any ***** ********* ****** *** ****** of * ***** ** ******* **** a ********* ** ****** *** ********.

*******, ******** ********** ********* ******** ***** and *****, ** ******* ** *** high (~**+) ******** ** ****** ********** of ***** ********, ***** ******* ** too *** (~** *** *****) ******** in ***** *********.

************, **** ***** **, ********** ******** (>30°) ** ***** ** ********* ** reducing ***** ****** (****** ***** *** light ** **** ** **) ******** in **** ****** *********** ***** *** is *** ***********.

******* *****, *********** *** ***** **** subjects ********** ********** ** **** ***** even ** ***** ****** (** ** 90° **********) *** ** *** ***** and ** ** ******** ****** (horizontal). ************, ***** *** *** ****** from *********** ******* ** ***** ********** on ********* ******* **** ** ***** facial *********** ********* ** **** ******, such *********************.

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

** *** *****, ***** ******** ******* hundred ****** ********, ********* **** *********, employees *** ******** ** ****** **********, as **** ** ******* ***** ** surveillance ******* ***** **** ******* ******* on *** ********.

*******, ***** ********** *** ********** ***** compared ** **** ************ **** *** is ****** ** ** ***********, **** as ******* **** ******* ******* ********, casinos **** ********* ** ********, *** city ******* **** **** ** ********* of ****** ********. ******* ** ****** populations *** ****, **** ***** *********** issues **********.

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

**** * ****** ******* * **** is ********, ********** ***** ************* **** ~80-90 ** ~**-**.

IPVM Image

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

******* **** ******* ********* ** **********, the ******** *** *** ******* *** false ********* (*** ******* ******** *** another) ********** * ***** ** *******, with * ********* ** **** *** subjects ** ******* ************.

IPVM Image

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

*******, ** ***** **** ***** ****** to ********* ***** ******** ******* ***** when ***** ***** *********** ********* ** 35.

IPVM Image

** ***** **** ******** *** ********* to ** ********* *** ****** ** recognitions ** ****** ******* ***** ***** not ****** *** ***** ************. ***** below **** ********* ********** ***** *********, with ******** ********** ********** ** ***** on *** *********.

IPVM Image

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

****** ~* ***** ** ******* ***** were ** ***** ************ ***** ****** suggested ********** ** **, **** ******** a ******* ** ****** *** ****** in * ********** ******, *********, *** outdoor ******* *** **** ******** ** subjects ******* *******.

IPVM Image

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

** * **** ***** ******* ******** were ********** **** * ********** ** ~80 - **, **** ********** ******* by ~** **** ****** ******** ******* of ~**' *** ********* ** ~**(*******).

IPVM Image

******** **** ********** *********** ** * 45 (******) ***** **** ~** **** confidence ** **** *****.

IPVM Image

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

********** ******* ** ~** ** * dark ***** (~*.** ***) **** **, resulting ** ************ ** ~*' ******** height **** ~**° ** ******** ***** the *********** ** ********** *********.

IPVM Image

******* ** **° ******** ** ***** confidence **** *** ** ********* *** missed ************ ******** **** ** *** mounting ******* ** ~*' *** ********* of ~**°.

IPVM Image

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

********** ** ************ *** ******* ** a *** ***** ***** ** ~** points ********* ** **** ************ ** lower ******** ******* ** ~*' *** downtilts ** ~**° **** ******* ****** the ******.

IPVM Image

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

********** ******* ** ~** **** *********** faces ** * *** *****. **** combined **** *****, **** ********* *** to ****** ************.

IPVM Image

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

******** ******* ********** ** *** ***** were ********** ** ~** ***** ********** drop **** * **** ** **** light.

IPVM Image

******* * *** ** *** ***** resulted ** ~** - ** ********** loss *** ***** ******* *********** ** higher ******** ******* ** ~**' *** downtilts ** ~**°.

IPVM Image

*******, * *********** ** ********** *** hats ******* ********** ** ~**, ********* in ****** ************ ***** ***** ******* thresholds.

IPVM Image

**********

***** ** *** ****** ** ****** once *********, ****** * ****** ** the ****** *** ** **** ** selecting * ********* ** *** ******** face ** ********* ** ***** (******* to * ******** *****) ** *** subject ** *** ******.

** *** ***** **** ** *** people ** *** ****** ** **** a ******** ********, *********** ***** ** be ****.

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

****** ******* ** ***** **** *** UI ** ***** *** ****** ** navigate. ****** ******* *** ** **** on *** **** ******, ** **** as ******* ********** *** ************.

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

**** * ****** ** ********** ***** is ** ***** ** *** ***** available, **** * ***** ***** ** the ****** *** *****.

IPVM Image

***** ** **** **** ** ** viewed ** * *** ***** *** such ** ********* ***** *****, ****** are ********** ** **** ** *** video **** *********** **** *** ***********.

IPVM Image

*******

***** *************** *** ** **** ** 1,950 € / ~$*,*** *** *** channel, ***** ******** * ******** **** of ** ** *,*** ********. ********* database ***** *** ********* ** ********** licenses. **** **** ** ********** ** slightly ***** **** ***** ****** *********** platforms ** **** ******, **** ** Anyvision (~$*,*** *** ****). *******, ***** emphasizes **** **** ******* ** **** a *********, ** **** ********* *** quoted *** *******, ********* ********* ** additional ********.

**** ******* **** **** *** ******* server ******** ******, ****** ******* $*,***+, depending ** ****. * ********* *** is ********, *** ***** ************ ****** by ****** *****. *** **** *******, a *** **** *** ** ****, an ********** $***-*** ***.

Versions ****

*** ********* ******** **** **** ****** testing.

  • *****: **.*.**

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

***** ** *** ***** **** ****** recognition ****** ** **** ****** ******** accurate ********* **** *****, ****** ** expect **** ********** ** **** ******* claims *** ** ********** **** *** during *** *****-** ********.

*******, ** **** **** ********** ********* *********, ********* ******* *****, *** ******* ********* ******** **** mask *********, *********:

Comments (3)

Is the licensing of about 1,950 € perpetual or per year?

BTW, IPVM guys, you should have asked Herta for the statistical curve of the algorithm, i.e. what FAR you were getting at what threshold value, as any FR vendor would have them based on their internal studies. Since your database was only 200 records, the stats tests as reported don't mean much. Somewhat depending on the scenario, the system operator would expect the system FAR to be order of say 0.01% (i.e. false positive in 1 out 1000 cases) for access control and higher than that to wake up security guard looking at a screen in the surveillance use case. Thus if you have a database of say 10K people, then the algorithm threshold should be in the order of 0.01%/10000 = 10E-7 (well, slightly larger than that, but that's a simple estimate).

Agree
Disagree
Informative: 2
Unhelpful
Funny

Those prices are perpetual. There's a 15% software maintenance charge per year, optional. Software upgrades are not included without the maintenance charge.

Regarding FAR, we have gotten it from multiple vendors, including Herta. However, we don't put much faith in it at all. Usually those statistics are developed using the NIST dataset, which uses much higher resolution images than typically provided in surveillance systems. For example, the NIST MEDS-II dataset is mainly mugshot images, so 200 pixel per foot images are not uncommon, and subjects are forced to look at the camera, which is much harder to enforce in the wild.

Compare a typical image from that dataset:

IPVM Image

To a typical, fairly high quality surveillance image:

IPVM Image

The subject is lower resolution (80-90 pixels across the face instead of 200) and angles are harsher.

Because face datasets I've seen don't reflect these realities, those figured developed using them are a rough idea more than a usable actual rate.

Agree: 1
Disagree
Informative: 2
Unhelpful
Funny

Thanks for the answer Ethan! If you dig into NIST reports you will see that the FRR is obviously a function of the database (i.e. the images like the one on top will have very low false rejects, but the images like the one at the bottom will have huge false rejects). However, if you set the threshold to say 20 then (for a good algorithm at least:)) the false accept rate will be the same regardless of image quality. NIST checks for that in some of their studies (the stability of the threshold vs. FAR is represented on the NIST curves as faint grey lines connecting ROC curves for different databases, which ideally should be vertical - none are perfect, but most will at least stay within the same order of magnitude from what vendors tell you).

Agree
Disagree
Informative: 1
Unhelpful
Funny
Subscribe to IPVM Research to read the full report.
Why do I need to subscribe?
The IPVM Research Service includes products tests and shootouts plus competitive and financial analysis, helping decision-makers better evaluate purchasing, partnering, developing, and/or competing against companies in physical security.
Already have an account?
Loading Related Reports