Hikvision DeepinView Camera Analytics Tested

By Ethan Ace and Rob Kilpatrick, Published May 08, 2019, 08:58am EDT

Hikvision is expanding its 'deep learning' offerings with a new camera series called 'DeepinView' claiming false alarm reduction and improved accuracy.

IPVM Image

We bought a DeepinView 1080p bullet camera (DS-2CD7A26G0-IZHS) and tested in multiple scenes for over two weeks to see how it performs examining:

  • False alert performance: What did and did not cause false alerts?
  • Intrusion detection accuracy: Did it miss valid people or vehicles in the scene in multiple conditions?
  • Face detection performance: Did face detection work reliably in full light and low light at multiple angles?
  • Ease of setup: How difficult or time-consuming was setup and optimization?
  • Vs. competitive analytics: How did DeepinView compare to camera analytics from our IP Camera Analytics Shootout - Avigilon, Axis, Bosch, Dahua, Hanwha, Hikvision, Uniview?
  • What components does it use? Ambarella, Huawei, Intel, NVIDIA, others?
  • What difference does firmware version make?

Performance *******

***** ** *** ***** of *******, ********* ********* of ******, *** ********** DS-2CD7A26G0-IZHS *** ******* *********:

  • ** ****** *********:****** ******* *** ****** managed ** ***** ***** alarms ** ******* ** the ***** **** ******* branches, *****, ***., ****** in **** ***** ******** cameras.
  • ** ****** ** ****: *** ********** ****** did *** ***** ** heavy **** ** *** scene, *** ** *** most ****** ******* ** false ****** ** *** past ********* *****.
  • ** ****** **********:****** *******/******* ******* *** scene **** *** ****** during *** ***** ** testing ** ******* **********, day, *****, *** ** rain.
  • **** ********* **** ******** in **** *****:** **** *** ******, face ********* *** *** miss *** *****, **** at ***** **° ******* or **** ********* ********.
  • ****** *************:************* ** *** ******** requires **** ******** ********* Detection ** **** *********, with *** ****** ********** to * **** ****** zone.

*******, ** ******** **** false ****** ** ******* sources *** ******* **** detection *********** ** *** light:

  • ***** ******* ******* ******** alerts:** *** *****, *** DeepinView ****** ********* ******* on ***** ******* *****, sheep, *** ****, * few ***** *** ***. Smaller ******* **** ** birds, ****, ********* **** alarmed ** *** **** frequently. ** ** ******* if *** ****** *** these ******* ** ***** or *******, ** ** does *** ******** *******.
  • ******* ***** *******: ******* ***** ******* such ** ****** ** a **** ******* ** or **** ********* ******** caused ***** ******.
  • **** ********* *********** ***** in ****/*** *****:** *** *****/**** ****** with ** **, **** detection *********** ****** ************, sometimes ********** ******* ********, especially **** ** ******* angles ** ******* ****** the *****.
  • ******** ******* *********: *** ********** ****** used ** ******* (****** through ** ************) ******* with ******** *.*.*, ***** performed ***** ******, **** false ****** ** *****, shadows, ****, ***. *** seen ** ***** ********. Hikvision ******** ** *.*.**, available ** ***** ******* but *** *** ** the *** ****, ****** they *** ** ** currently ** *** ******* of ***** ********* *** added. *** **** ******* are ***** ** *.*.** and ** ********* *** users **** ******* ******** to ******* ** ****.

Vs. ******** ****-******** *********

********** *** ********* ********* by ***** ***** ******* which ********'* ** ****-******** analytic ******* ********. ** our *****, ** ******** camera ** *** **** scenes *** *** ********* by *** ******* ****** several **** ** *******, nor ***** *******, *******, foliage, ***.

Vs. ***** ********* *********/***

********** *********** ** *** tests ** ****** **** of **** *********** ********* ** **** tested, ********* **** (********* Defender *** ****** *****), Hanwha ******* *, *****, and *********'* *** ******** intrusion ********* *********, **** much ****** ********* ** foliage, *******, ****, *** even ***** *******.

***** **** ********, *** only ******** **** *** shootout ******** *********** ****** DeepinView ** *****'* ***, which ******** *******, ****, foliage, *** ****, ****** occasionally *** ****** **** shadows. ***** ***** ** more ******* **** *********, however, ********* ******* ***** to ********* *** ******* calibration *** **** ******, instead ** ****** ******** the ****.

******** ** ******** ***, performance ** *********** ******, as **** ****** **** VMD ***** ****** ***** on *** ***** ******, including *** *** ******* to *** ****** ***** alert ******* *****, ******* discriminating ******* ******, ********, or ***** *******. *******, DeepInView ** *** **** expensive **** ************ *******.

*******

*** **-*********-**** ***** *** ~$900 *** ******. **** pricing ** ********** *******'* ********* **** **(******* ***** *** *****/******** camera), ****** ********** ****** than********'* ******** ******* (~$***-*** ********* street *****).

*** ***-** * ****** equivalent ** **** ****** (**-*********-****) ***** *** ~$*** online, ****** **** ******** a ~$*** *** **.

Physical ********

*** ********** ****** ** physically ******* ** *********'* non-deep ******** ****** ******. It ** ******* ** size *** ************ ** other **** **** *******, such ** ****** (***** below), ******** ******, **** an ******** **** ***.

IPVM Image

** ****** *** ******** of *** ****** ** this *****:

DeepinView ********

*********** *** ********** ****** is ******. ***** *********** only **** ** ****** Intrusion ********* (** **** Detection). ***** ******** *** be **** ** ******* (full ****** ********* ****, and *******/******* ****** **** settings), *** *** * limited ****** ** *********** in *** *****.

** ****** *** ***** setup ** **** *****:

Rejected ***** ******: ****, *****, *******

*** ********** ****** ******** many ***** ***** ******* that ******** *********** ** past ********* *****.

*** *******, ********** *** not ***** ** ***** rain ****** *** *******, through ******* ****** *** and *****. **** *** one ** *** **** common ******* ** ****** in **** *****, **** nearly ***** ******** ****** frequently ******** ** *********.

No-False-Alerts-In-Heavy-Rain-Wind

*** ****** **** ******** alerts ****** ** ******* branches/foliage ** **** ** the ******* **** ******, another ****** ***** ** past *****.

No False Alerts On Brush Shadows

Regular ***** ****** ** *******

*** ********** ****** *** notable ****** **** ***** alerts ** ***** *******, such ** **** ** the ******** ** *** field ** **** *****.

False-Alert-On-Dogs-In-FOV

** **** ******* ** goats *** ***** (***** below). ** ** ******* whether *** ****** ********** these ******* ** ** people ** ********, ** it **** *** ******* differentiate ** ******* ****** of ****** ********.

False-Alert-On-Sheep-In-FOV

Fewer ****** **** ***** *******

**** **** *** ****** also ************ ******* ** small *******, ********* * few ***** ** *** testing, ****** ** * much ***** **** **** larger *******, ***** ********* nearly ***** **** **** were ** *** *****.

False-Alert-On-Birds-In-FOV

*******, ** ******* *** cat ** *** ***** below, ***** ***** ******** on **** ******* ****.

No-Alert-On-Cat-In-Scene

Alarms ** ***** *******

******* ****** ****** ** false ****** *** ******* light *******. **** ******* in ******** ******, ********* outdoors ****** * ************:

Alert After Lightning Flash

No ****** **********

****** * **** ** testing ** ******* ******, people ** **** **** not ****** **** ******** the ***** ** **** of *** ****** ******* or *******.

No-Missed-Detections

****** ********* **** ********* well ** ***** ****, alerting ** *** ****** walking ** *** ***** and ******** *** ****.

People Detection In Heavy Rain

Person/Vehicle ********* ***** ***********

** *** *****, *** DeepinView ****** *** **** to ******** ****** ****** at ***** ~** *** (~160' ******** ***** * ~175' ****). **** **** detected ** **** ** this ********, ~***', ***** 3-4 ***.

**** ** ********** ****** than ******** ****** *** day (~* ***), ****** lower ** *****, ***** Avigilon ******** ** **** as ~** *** *** detection. ***** *** ******** day *** ***** ** ~4-6 *** ** *** tests.

IPVM Image

Face ********* ***********

*** ********** ****** **** includes **** *********, ***** may ** ******* ******* of ********* *********, *** the *** ****** ** used **************. **** ******* is ******* ** ********* to ** ** *** "VCA ********" *** ** basic *************. "***** *****" refers ** ******* ******** rules **** ** ********* detection *** **** ********.

Smart-Event-Or-Face-Capture,-Not-Both

** *** ******* **** detection ****** **** ** a **** ***** ***** (~200lx), **** ********* *** subjects **** ** ******* degrees ** **** (~** degrees) *** ******* ******** across *** ***** ******** only **** ** *** face.

Face-Detection-Works-Well-In-Full-Light-Harsh-Angles

** *** *****, **** IR ** ** ***, performance *********, **** *** camera ******* ***** ** subjects ******* ****** *** scene.

Face-Detection-Misses-In-Low-Light-IR

Movidius ****** * *** ****

*** ********** ****** **** an***** ******** ****** * vision ********** **** (***), ******** *****.

IPVM Image

******** ** ************* ***, ***** **** * single ***** ****** ********* for ** ******** ** intrusion, ********** * ****** 2 ** * ****** channel ** ***** ******** much ******* ********** *****. As ******* **** ** the ***** ****** * in *** **** *** years, ********** ***** *** likely, ** *** ****** X ** ******* **** powerful **** *** ****** 2 (*** *** ****).

Unknown ***

** **** **** * private ******* ***, *** the ********* ** ********* labeled ***** ********* **** in ***** *******:

IPVM Image

* ****** *********** ****** suggested * ******** *******:

  • * *********-******** ***, ******* of ***** *** ***** party ***** **** **** historically **** ** ***** models. "**** ***** **** them *** ***** ******** they **** ** * device ** ******* ******** pricing ****** ********."
  • * ********* *** **** another *********, *.*., *********. Given ***** ********** *****, Hikvision *** ** ***** chips ** ****** **** their **** ** ****. "It **** *** **** that **** ****** ** order ***** **** **** your *** ****** ******** on ****... ** ** could ** **** * normal *** *** ***** HiSilicon ****, *** **** Hikvision ********. "

Versions ****

*.*.** ******** *** **** for *** ********* **-*********-**** and ** *** ***** of *** ********.

** **** ****** *.*.* and *.*.** (**** *** shipped ** ** *** is ***** ****** ** Hikvision *** ** *******), accuracy ** ***** ******** was ******* *****, ********** 5.5.5, ***** ******* ** shadows, *****, ****, *** more ******* ***** ***** versions *** ***. ** did *** ******* *** examples **** *.*.* ** 5.5.81, *** ***** ****** be ***** ** *** substantial ********** ** ***********.

Country ** ******

*** **-*********-**** *** **** in *****.

Comments (15)

Any data on range/FOV? Eg was the stated performance seen at a range of 100M over a 90degree FOV, or 20M over a 30degree FOV?

Agree
Disagree
Informative
Unhelpful
Funny

In our tests, the DeepinView camera was able to reliably detect people at about ~11 PPF (~160' distance using a ~175' HFoV). Cars were detected at over 3x this distance, ~530', about 3-4 PPF. Those numbers are using a ~60° FOV.

Agree
Disagree
Informative: 2
Unhelpful
Funny

Here is some backstory to the test.

Hikvision USA tech support was not helpful with this model (not blaming them, this is a common pattern we see, they are generally not trained on new / advanced products). Tech support verified what was publicly shown in the US that .81 was the latest.

This time, Hikvision management actually provided tech support. We reached out to European contacts who confirmed that .91 was better. To Hikvision's credit, the USA team gave us a direct link to the .91 firmware so we could test.

You will recall last year's disastrous DeepInMind NVR test where Hikvision refused support.

Related to that, we have updated firmware for DeepInMind. However, the USA is only selling the 'A' hardware version currently (which is what we bought) while Europe has the 'B' hardware, which is presumably better. We are trying to navigate through this.

P.S. - please Hikvision, don't ship until you have fully optimized firmware. If you still need to improve, that is what beta is for, production customers (at least in the USA) expect production-ready products.

Agree
Disagree
Informative: 1
Unhelpful
Funny

Letting the market test products is common for Chinese manufacturers. In my experience with both dahua and hikvision virtually all first release products have a huge range of bugs and require a a lot of adjustments to the firmware to get them right. Most of the information they use comes from distributors or system integrators who have had the misfortune of buying and installing these types of first release products which  really are just a “beta” style product.

Agree: 2
Disagree
Informative: 2
Unhelpful
Funny

But .... where did you get the sheep from ?

Agree
Disagree
Informative
Unhelpful
Funny: 7

The issue with the sheep goes into GDPR territory. Did anyone seek their permission? Has AI been used without the birds, cats or sheeps knowledge? What's been done with the stored data and do any of the targets form part of an ethnic minority group? I'm not convinced IPVM is taking their data protection responsibilities seriously.

Otherwise a very decent report Ethan. I'm curious though - during the two-week test just how much learning took place and did the false positives improve over time - giving the potential for further improvement?

And slightly off-tangent, given Trumps latest sabre-rattling - how do tariffs affect the Intel product element of the camera? Just curious.....

Agree
Disagree
Informative: 1
Unhelpful
Funny: 1

What's been done with the stored data and do any of the targets form part of an ethnic minority group?

I think you are joking but I don't think it's funny to make use of human rights abuses. Accept that Hikvision is profiting off it, also see The Chinese Surveillance State (Part 1 And 2)

As for Intel, tariffs would not be much of an impact but sanctions against Dahua and Hikvision for their profiting from human rights abuses would.

Agree
Disagree
Informative
Unhelpful
Funny

I put that in there for you John so you could have your fix.....

Disappointing you didn't engage with the technical query though, and just took the opportunity to get the soapbox out.

Agree: 2
Disagree
Informative
Unhelpful
Funny: 3

I did not administer the test. That you think it's funny to joke about human rights abuses says a lot about you.

I'll have Ethan address your question about improvement. In the future, if you want to ask a technical question, ask it without joking about oppressing ethnic minorities.

Agree
Disagree
Informative
Unhelpful: 3
Funny

I won’t seek to undermine or devalue Ethan’s work by perpetuating your rant and detracting from the report.

Agree: 2
Disagree
Informative
Unhelpful
Funny

You have undermined and devalued Hikvision by bringing up their profiting from human rights abuses in a post that had nothing to do with that.

Agree
Disagree: 1
Informative
Unhelpful
Funny

We saw no improvement over time. It’s pretrained, and may improve from firmware version to firmware version, but it did not stop alerting on branches after several days in the same scene or start ignoring dogs after seeing them in the same area. Also there is no way to mark events as valid or false, so it can’t be manually trained. 

Agree
Disagree
Informative: 2
Unhelpful
Funny

Thanks Ethan.

Agree
Disagree
Informative
Unhelpful
Funny

the unknown SoC should be  "Huawei Hisilicon 3519 V101" 

Agree
Disagree
Informative: 2
Unhelpful
Funny

#4, thanks. How did you figure that out? Is there some pattern or practice that allowed you to identify a specific Hisilicon chip?

Agree: 1
Disagree
Informative
Unhelpful
Funny
Read this IPVM report for free.

This article is part of IPVM's 7,201 reports and 959 tests and is only available to subscribers. To get a one-time preview of our work, enter your work email to access the full article.

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