Verkada Below Average Facial Recognition Tested
Verkada told the Washington Post "its facial recognition system accurately matched more than 99 percent of faces". But after multiple calls between IPVM and Verkada's engineering team, and weeks of us testing Verkada, we found the accuracy far worse and inferior to most competitors we tested.
IPVM tested Verkada's facial recognition, and inside this report, we detail the full results of our testing, including:
- Strengths and Weaknesses found during testing.
- Facial recognition with shallow, moderate, and harsh camera down-tilt angles in full light (~180 lux).
- Facial recognition in low light (~1 lux) with IR.
- Facial recognition with sunglasses, hats, and masks added.
- Identification of people within groups.
- Face "De-Duplication" testing.
- IPVM stress testing, showing enrollment, total matches, and false match rates.
- Interface overview of the system.
- Bulk uploading of faces for watchlists.
- Approximate PPF needed for facial recognition.
- Pricing and Availability.
This report expands upon the Facial Recognition Rankings 2023 - Avigilon, Briefcam, Hikvision, IronYun, Oosto, Rhombus, Turing, and Verkada, by providing more details of Verkada facial recognition performance.
************
*******'* ****** *********** ** *********, ** no ********** **** **** ****** ******** licenses, ** ******* ******* *** *********. However, *******'* ****** ************* *** ********* ** *****, ********, and ********, **.
Command ********* ********
******* ******* ********* *** *** **** ("People" *** "*****") ***** ***** *** utilize ****** ***********. ** *** "******" tab ***** * ******, ***** *** click ** *** ***** **** ***** a ********* *** ****** "****** ** this ******" ** **** ******* ********. Alternatively, ***** *** *** *** "*****" tab *** ***** ** *** **** thumbnails ** **** ******* ********. ***** is * ***** ***** ** *** Verkada ******* *********.
******* ********* ***** ***** *** **** up ** ** ******* ** ****** on *** "*****" ***, *** *** all ***** *** ** ********.
******* **** ****, *** *** **** accurate ******* ***** **** *** *** "People" *** *** ******* ****** ***********.
Verkada ******** *********** ******* ****** ** ******** *** ********
******* **** **** *** ***** ***** to ***** ****** **** ** ** marked ** * "****** ** ********" for **** ************* ********* ******. ******* said, ***** * ******* ** ** entry ***** *** **** ******, *** allows *** ****** *** ******** * better ************** ** ***** *******.
* ******* *** ****** ** ******** (POI) *** ********* ******** ** *** products. * ******* ** * ***** image ** * ******’* **** **** a ******** *** *** ** ** entry ***** *** ******. * *** is * **** **** * ******** has ******** ** ***** ** ******** an *****.
******* ***** **** **** ****** ****** precision *** ****** **** *** *************, which *** ******* ** * ****** alert *** ******* *** ****** ****** a **-****** ****** **** *** ********, using *** "**** ******" *******.
**** *** *******, ** ***** ****** higher ********* *** ****** **** *** notifications **. **** ******.
****** *******, **** *** *** ******* a ********** ******* ******** ****** ** POI ** ******. *** **** ************** observed *** *** ******* ** ******* alert ************* ********** **** *** ****.
** *** ******** *******, ** ******* the ******* *** ***** ******** ******** of *** *** ***** ******* ****** recognition ****** *** ***** ** *** not.
Executive *******
***** ******* ****** ** **** "**** than ** *******" ****** *********** ********, IPVM ******* ***** *** ******** ** significantly ***** **** *** ******* **%, matching **% ** *** ***** ** ideal ********** (***** ******* *****, ****-*** scenes), ***** **** ******** **** **** 15% ** *** ***** ** ***-***** conditions (******* *****,***-*****/**** ******), *** ** exhibited ******** *********** **** ***** **** it * **** *** *** **** use *****.
*** **** *********** *********** ********* ** that ** ** ********, ** ** additional ****, **** ******/*********** ** *******, so **** *****, **** ***-*********** ************, may ** ********* ***** **** **** are *** ****** ***** *** **.
Technical **** *******
**** ******** ** ***** ****** *********** offerings *******'* *********** ** ******** *** to ***-******* ** ***** ******* *** partial *****, *** ******* ** ********* identify *** ******** ** * *****. Additionally, ******* ******** ** ******* ** faces ** ***-*****/**** ****** **** ** on *** **** *********** ***** (****, sunglasses, *** *****) *** **** ***** rate *** ***.
** *** ******** ****, ******* ******** quick *************, ********* ****** ~* ******* via *** ******* ********* *** *****, however, *** ****** *** ******* ** a ****** ***** *** ******* *** camera ****** * **-****** **** *****. Verkada's ****** *********** ** **** ** set ** *** ****** ***** ** programmatically *** ******-**-******** ******** ******* *** API.
*********
- **** *****:******* ****** *********** ** **** ** set **, **** **** **** * clicks ****** *** **** ****** (****** recognition ** ******** ** *******).
- ***** ******:******* ******** ***** ******-**-******** ****** (~* seconds) ******* *** ********* *** *****.
- ************ *********:******* ****** *** ************ ********* ** people, ** "****** ** ********" *** Verkada's ***.
**********
- **** ******** *******:******* *** * *** ******* ***** rate *** **** *** ******* ***** in ******** *** ***** ****-**** ****** in **** ***** (~*** ***). ************, with *********** ***** ******** ** ******** degraded.
- ** ***** ** ***-*****/**** ******:******* ******** * *** ******* ***** rate **** *********** (****, **********, *****) in **** ***** (~*** ***).
- ************ ******** ** ******:******* *** ************ **** *** ************** of *********** ****** ******.
- ***** ******* ********:******* ******** ******* ***** ** *** "Faces" *** ** ********* ***** ***** across *******, **** ** ****** *** users ** ***/****** *** ***** *** to **** **** **** "**-***********".
- **-****** "****-****":"****** ** ********" ****** *** ******* to * ****** ************ ***** **** per ******, *** ****** ** * 15-minute **** ******.
Simple ****** *********** *************
******** *******'* ****** *********** ** ******, and ******** * ****** *** ****** overall ** *** ******* ******* ********* to ** *******, ***** *****.
** ********, ***********, **** *********** ********** (*** ****), ***** ******** ******* *************, *** additional ********, **** ** ** ***, whereas ******* ** ****** ** *** cloud, *** ******** * *** ****** per ****** *** *************.
Poor *********** ** ******* ****** ****-**** ******
****** *******, ******* ********* ****** **** cameras ********* ** ******* ****-**** ****** (7' - **°) ** **** ***** (~180 ***), ******** ***** *********** *** camera ******** ~**% ** *** *****, and **** ~**% **** ******** *********** at **°.
***** ******* ******* ~**% ** *** faces *********, ** ****** ** ******* the ******* ~**% ** *** ***** and *** *** ***** ~**% ** the ***** ** *** *********.
*** ******* ***** (~**° ********) ** the *****, ******* *** * ******* face ***** **** ** ~**%, *** not ***** ~**% ** *** ***** to *** *********, *** ******** * single ***** **** ***** ***** **** failing ** ****** * ****** ****** a ****** ****.
*** ******* ***** (~**°) ** *** scene, ******* *** *** ******* *** correct **** *******, ****** ** ****** a ****** ** ~**% ** *** walks, *** ******** * ***** **** match.
Poor *********** ** ******** *** ***** ****** ****-**** ******
****** *******, ******* ********* ****** **** cameras ********* ** ******** (*' - 22°) *** ***** (**' - **°) down-tilt ****** ** **** ***** (~*** lux), ******** ***** *********** *** ****** directly **** **** **% ** *** walks, *** ******** ******* ***** ~**% of *** *********.
******** ****-****:
***** ****-****:
Poor *********** **** ***********
****** *******, ******* ********* ****** ** full ***** (~*** ***) **** ******** wearing *********** (****, **********), *** *** less **** ~%** ******* ******* **** subjects *********** *** ****** ********. ************, with ******** ******* ********** *** *****, Verkada **** ******* * ****** **** correctly *** ** ** *********.
Poor *********** **** **** *****
****** *******, ******* ********* ****** ** full ***** (~*** ***) **** *** subject **** **** *****, ********* ******** faces *********** *** ****** ******** **** than **% ** *** *****.
Poor *********** ** *** *****/**** ** **
****** *******, ******* ********* ****** ** low-light/Dark ****** **** ** **, *** matching *** ***** *** ** ** walks *********.
Poor *********** **** ******
****** *******, ******* ********* ****** ** full ***** (~*** ***) ** ******* Down-tilt ****** (*' - **°) *** groups, ********* ******** *** *** ** 6 *****.
************, ******* ******** ***** *** ** matches ** **** ** *** ******** within * ***** **** *** ******** were ******* ** * ****, ********* one ******* ** * ****. ********, the ******* "***** ****" ****** **** multiple ****** ** * ***, *** person ** ***** *** ****** *********** is ***** ** *** ******* ********* below.
Person ** ******** ******
******* ****** ***** ** *** ** "Person ** ********" ******, *** **** the ***** ** *** ****** ***.
*** ***** ****** *** **** *** delivered ** ~* *******, *** ******* the ********* *****, ****** **** *** image, ****** ****, *** ****.
******* **** ******** *** ****** ** its ******* *********, ******* *** ********* video *** ***** ******* ******** ** the *****.
******* **** ****, **** ** *** perform ************* *** "****** ** ********" notifications, *** ***** ** * **-****** cooldown ****** *** ************* ** *** same ****** ** *** **** ******.
*** ****** ** ******** (***) *************, two ********* ****** ** ****: (*) we ** *** ******* ******** *** to *** ******* ************ *** *** notifications. (*) *** ***************, *******, **** * ** ****** cooldown ****** *** ************* ** *** same ****** ** *** **** ******. We ******** **** ******** ****** ** avoid ******* ********** ************* ** *** customers *** * ********** *** ****** a ***** ******.
**** ******* *****, ***** ** * 15-minute ******** ****** ******* *** ******.
***** *** *** ****** *** *** or ******** (** ** **) **** on *** ****** ***, *** ****** to ** ******** ******* *** *********, email, ** ***.
Faces ******* ** "*****" ***
****** *******, ******* ******** ********** ****** detection, *******, **** **-**% ** ***** were ***** ** *** *********. ******* told ****, ***** ****** ** ******** to **-********* **** **********, *** **** expect ** *** ***** "****" ********** than "******" ********* ** *** *********.
** ******** *** ********* *** ****** to *********** **** **********. *** ****** page *** *** ********* ****** ******* and *** **** **** ** ** entry ***** *** ******. *** **** reason, *** **** *** **** “******” detections **** “****” **********.
******* ********* **** **-*********** ***** ** batches, *** ********* ** ******* ********* user *******, *** **** ** ***, the ***** **** ***** ****** **** a *** ******* ** *** ****.
****’* *** *** **** ************* *****: our ****** ********* ***** ** *******, and *** **** ** **** ***** depends ** **** ******* (*.*., **** traffic, **** ** ***, ***.). ***** the ******* **** ** **** ***** for ************* ** *** ** *** hour, **** ******* **** **** * few *******. **** ******* ** ** perform *** ******** **** ** ******* to *** ********* ******* *********** *** much ******* ** *** **** ****** pipeline.
******* ***** **** ** *** ******* de-duplication *** *** ****** ** ******** notifications, *** **** **** * **-****** cooldown *** ************ ** *** **** person ** *** **** ******.
*** ****** ** ******** (***) *************, two ********* ****** ** ****: (*) we ** *** ******* ******** *** to *** ******* ************ *** *** notifications. (*) *** ***************, *******, **** * ** ****** cooldown ****** *** ************* ** *** same ****** ** *** **** ******. We ******** **** ******** ****** ** avoid ******* ********** ************* ** *** customers *** * ********** *** ****** a ***** ******.
Testing **-*********** *** *****
**** ********* ** ***** ****** *** camera ** * ****-*** ***** (~*** lux) **** *** ****** ******* ** a ******* ****-**** (*' - **°) but ******* *** ** ***** * minutes ******* **** ****.
*** **** ******** ***** ** *** "Faces" *** *** ************ ****** ******** walks, ******* **** ** ***** ** three ***** *** ****, ***** ** the *********. **** *** *** ******* any ***** ******* ** *******'* ******* "De-Duplication", *** ** ****** **** ********/******.
Does *** **** **** ******, ****** ************ ********* **** ***
******* ****** ***** ** *** ***** one ** *** ** * ********* using ***** ******* *********. ** *** a **** ** * ****** ** interest, ***** **** ******** ** *** "People" ***, **** ** ** "********" and ***** "****** *******." (***** *****) Then ***** *** ****** ****** *** exceeding ** ** ** ****, ****** a ******* ** * ****** ** interest, *** *** **** ** *** profile.
*************, ******* ****** ***** ** ******, update, *** ****** "****** ** ********" using ***** ***. ** ****** * face ***** ***** *** ***, *** users **** ******* *** ***** ** base64, *** **** * ****** ***** can ** ******** ** * ****.
****** **** *******, ** **** * script ** ************* ****** "****** ** Interest" ******** ** **** *** **** ~2,500 ******.
*** ******* ******* ********* ************* ******* images **** ** *** **** ***** specifications *** ******** ** ***** ******* on *** ******* ******. ** * face ** ********** ********, ** ***** in *** ********* ***** *** "********", and "****** ** ********" ** ********* by * **** **** ** *** top **** ****** ** *** *****.
** ********, **** *********** (****** *** IronYun) ** *** **** ****-****** **** uploading, *** **** ***** ******* ***** to ******* *** ********* ******** ** request **** ** * ***** ********* is ********/******.
IPVM ****** *******
**** ********* ****** ******* ********* ***-* *************, ******** *** ********** **** ** faces, ***** ***** ****, *** ***** match ****. ** **** ~*,*** ***** to ********* *** ********** **** ** the ******, ***** ******** ~*,*** ********** face ******.
**** *** ********* *** *******, ** recorded * ********* ****** ******* *** videos ** ****** ******* **** **** and ******* ******, *** ******* **** detected ~** ****** **** * ******* face ******* *** * ***** **** matches.
~120 *** ******** *** ******** ****** ***********
****** **** *******, ******* ******** ~*** PPF ** **** *** ******** ****** recognition *** ******* ********* ** ******* down-tilt ****** ** **** ***** (~*** lux).
***** *** *** ******* ******* ** the ****** ******** *** *** *****, this ****** ********* **** ** ****** for *******'* ****** ***********.
Firmware ******
******* ****-*: "** ** ****"