Hikvision Launching Deep Learning Recorders

Published Jul 20, 2017 14:27 PM

Hikvision has become a common choice for super low cost NVRs.

Now, Hikvision is aiming to move up market, with deep learning NVRs that claim far greater intelligence and accuracy. In this note, we examine the line, share feedback from Hikvision and look at its potential market impact.

Deep ******** ****

*** ****** ** ***** ****** **********, with *** ***** ******** *** ***-*******-**/*** [**** no ****** *********], * *** ********* supporting ** ** ** ******** ********* and ** ******** ** "***** **** Detection *** ********".

******* **** ****** *****, *** *********'* ***** * **** range ** **** ******** *** ********* capabilities:

*** ******* ** ** ********** ******, a ******* ******* ********* **** ********'* recent *******. ***** ** ** ******* from *********'* ****** ***** **. ****** it ** *** ***** ** ********* from **********, *** **** ********* ***** seems *******:

Release ****

********* **** *** **** ** ** release the ***-*******-**/*** **** ***** ********* ** *** end ** ******.

Claimed ********

****** ********* ****** ** ******** (**%+), Hikvision ********** **** ******** **** **** data *** ********** ******** ********. **** is ********* ******** **** **** ********, unlike ****** *********, ** *** ********** can ******* **** **** ******** *** more ******** **** ***********.

Update: **** ******** ** * ******

*********'*********** ******** ******** **** *** **** ******** ******* will ******* * ***** ***** ****** to ************ ******, ** ************ ** the ********* *****:

 

************, * ***** ******* ** ***** alerts ** **** ********* (** ***** to *****) ***** ****** ***** ** triggered ** ****** ** **** ** wind ** *******, ***. **** ********* is ******** ** ** **** ** to **** ***** ****** *** *** if *** ****** ******** *** ****** using **** ********. **** ***** ************* improve ***********.

*******

********* **** **** ******* *** *** been *** *** *** **** ** will ** * ******* **** * standard ***. *** **** ** * premium **** ** ** *********** ******, as *** **** ********* ** ******* deep ******** ** ******* ****** *** the ********* ***** **** **** ***** provide (** ********) ***** ** ** well.

Vs ********* **** ******** ******

********* *** **** ************* **** ******** cameras *** ****** * ****, ********* an ******** ****** ******** ******. *** structural ********* ** **** ********* * deep ******** *** ** ** **** it ****** ** *** ******** / conventional *******, ****** **** ****** *********** ones.

Updated: ****** *******

**** ******, *** ********* **** ******* streaming ***** **** ********* **** *** ***** cameras *** *****. *********, *** ********* **** only ******* ***** **** ********* **** using Hikvision ***** *** ***** *** ******. On ****, ********* ***** "*** *** function ****** ** ******* ** *** camera. *** ***** **** ********* ** done ** *** ***." ****** ******** upgrade **** *** **** ******** ******* for *** ***** *******.

LTS '***'

*** ******* ******* ** **** ** that********* *** / ********* *********** ***** ******* ********* **** **** *** video *****:

***** *********'* *** ** ***** *** own ***** **** *** ******* ****, LTS' **** ********** ****. ** ********* wants ** ***** * ********** *** premium ******** *** ***** ********* *** price ************, ** ***** ** ****** to **** **** ******** ********* ** Hikvision *** * ******* ******* ******.

Potential ******

** **** *** ***** ******** *** more **** * ******, *** *** question ** *** ******** / **** will ***** ********* ****. ** *** plus ****, ********* ** ***** * newer ********** ** **********.

*** **** **** ******** *** ********* is **** **** ******** ********* ***** well *** **** ***** ** ***, which ***** ****** *** ***** *********** looking ** ************* **** ***** *********.

*** ***** **** ******** *** ********* is **** *** **** ******** ********* work ****** *** ****** *** ********, leaving **** ******* ** **** ***** manufacturers *** **** ********* **** **** the *****.

****** ***, **** ********* ** ************ moving ** ******* **** ******** ** a ****** ** *** ****** **** they **** ** **** **** *********'* expansion **** **** ********* ********* *** segment. **** ***** ** **** *** DeepInmind **** **** ********.

Comments (39)
Avatar
Brandon Knutson
Jul 20, 2017
IPVMU Certified

This is exciting as it will push the innovations of other VMS providers. On the downside I'm not comfortable putting overachieving Hikvision appliances on my network. 

(4)
(1)
(1)
(3)
UI
Undisclosed Integrator #1
Jul 20, 2017

Now they(Hik) will have access to the network and intelligence inside it!

(5)
(1)
(1)
UM
Undisclosed Manufacturer #2
Jul 20, 2017

Perfect product for Hikvision, now the recorder can notify them when there is activity of interest at thier customers' sites and they don't need to have so many spies back in China manually evaluating as much video.  😀

(3)
(1)
(1)
(6)
GF
Giancarlo Favero
Jul 21, 2017

I think that, as the Datagate scandal disclosed, with the americans intercepting for years the fixed and cellular phones of Angela Merkel, Silvio Berlusconi and thousands of other Presidents, Prime Ministers and personalities all over the world, as far as spying others is concerned, the americans should be the first to SHUT UP, if I may (and of course can) speak candidly.

Not to speak of the NSA agents opening the packaging of PCs and servers at the Customs, installing every sort of malware and remote control programs, and repacking everything.

Come on wake up !

Giancarlo Favero

(5)
(1)
UE
Undisclosed End User #5
Jul 21, 2017

Right! Why go through the trouble of installing spy software when it can come pre-installed from the factory!!

(1)
(1)
(2)
(2)
U
Undisclosed #6
Jul 22, 2017

you just made the TSA no-fly list.  :)

(2)
RK
Rashid Khan
Jul 20, 2017

This is no doubt exciting and positive for the industry. This will also create healthy competition with AI / Video Analytics / Facial Recognition industry heavy weights who were dragging the industry down by milking profits as much as possible until now where they will be forced to bring their costs down introducing products to the lower segments of the market creating larger adoption rates. All in all good for everyone, lets see when Dahua are releasing similar products, I know for a fact that Dahua already have and have released such products in Asian markets.

(2)
(2)
JH
John Honovich
Jul 20, 2017
IPVM

AI / Video Analytics / Facial Recognition industry heavy weights who were dragging the industry down by milking profits as much as possible

Those guys only wished they were milking profits. Almost all struggle to stay alive.

I don't know how Hikvision will do here but the perception that the existing analytics players were generally profitable / strong is simply false.

(4)
(3)
JH
John Honovich
Jul 20, 2017
IPVM

To add to that, I am curious how much and what level of field services Hikvision (and Dahua) will provide for these deep learning systems. A big part of traditional video analytic costs is having experts on-site at deployments optimizing systems. So it might be that their deep learning ones work so good that it no longer needs it but, if that is not the case, either they will face unhappy customers or high local costs to deal with issues.

(2)
RK
Rashid Khan
Jul 21, 2017

The whole point of Deep Learning & AI is that it is self learning and self adapting, requiring minimal configuration, for example it is not rules based but object based, which by itself removes a lot of false alerts and simplifies a lot of stuff like skyline, horizon, sun, cloud, lighting changes etc. Of course it will still require experts but the man hours should be a lot less than for traditional video analytics.

(1)
Avatar
Brian Karas
Jul 21, 2017
IPVM

The whole point of Deep Learning & AI is that it is self learning and self adapting, requiring minimal configuration, for example it is not rules based but object based

That is not universally true. Deep learning concepts can be applied to analyzing activity (which BRS Labs tried and failed at), but is more commonly applied to the object detection/classification part of analytics. After an object is classified the system needs to determine if that object is in an area, or doing some action (loitering, moving the wrong way, etc.) that should generate an alarm.

The alarm-generation side is still primarily rules-based, though several companies are working on and previewing automatic anomaly detection analytics. Still, it is far from common and not a given when "deep learning" is involved.

Additionally, while deep learning is often being used to imply reduced or eliminated need for calibration, there can still be cases where manual intervention is needed as part of the setup or optimization process to get optimal performance.

And even if/when you do have a fully automatic "touchless installation", there is still a lot of customer training involved to teach them how manage alarms, how to understand how/why the system made particular decisions (or missed events).

Any company trying to sell analytics without understanding the heavy sales load in even the best case scenario is unlikely to have broad success.

UM
Undisclosed Manufacturer #9
Jul 25, 2017

"Any company trying to sell analytics without understanding the heavy sales load in even the best case scenario is unlikely to have broad success."

This would hold true for most developers of these systems, however if HIK shift hardware at their current low cost and offer the analytics as part package, then there will be a lot of installers and end users tuning and tweaking to add accuracy to their searches.  If all that data accuraccy can be extracted then it provides foundation to improve the code at source, ultimatly leading to a smarter out of box solution, while limiting the ground level support required due to the free/reduced risk of using the solution, just as Google and others collate their big data sets. 

As for cyber security and HIK getting the data returned, as HIK will continue to sell high box numbers on cost any failure returns will certainly provide a good amount of returned data without a second thought being given by end users and installers.  It is after all not just network connected devices that pose risk of data capture.    Fault, Returns number, gone!  Complete with any network config and user passwords, etc.

 

 

JH
John Honovich
Jul 25, 2017
IPVM

however if HIK shift hardware at their current low cost and offer the analytics as part package, then there will be a lot of installers and end users tuning and tweaking to add accuracy to their searches

We will see my gut feel would be the opposite. People who buy low cost products through distribution are generally looking for reliable, straightforward things to install and move on. If they have to go back and forth to 'tune' and 'tweak' the system, I do not think most of them will view that as a positive as that is operationally disruptive and costly.

Avatar
Brian Karas
Jul 25, 2017
IPVM

"tuning and tweaking to add accuracy to their searches. If all that data accuraccy can be extracted"

This would depend on how that "tuning and tweaking" is carried out. In essence, users would need to be specifically flagging false alarms and valid alarms to send useful training data back to Hikvision. If they are just adjusting coarse settings like sensitivity, delay time, etc., that would not be very useful data.

 

 

RK
Rashid Khan
Jul 21, 2017

I know for a fact that the big players and I am talking about NEC & OT-Morpho (Safran) were making lots of profit and i know that because I have actually worked for them. The profit margins were simply sky high! Now they will face a little bit of competition and struggle. This is what i meant.

(1)
JH
John Honovich
Jul 22, 2017
IPVM

I am talking about NEC & OT-Morpho (Safran)... Now they will face a little bit of competition and struggle.

Rashid, your comparison is unfair to Hikvision. Hikvision is nowhere claiming that they are competing with the likes of NEC and OT-Morpho who do biometric projects for governments, etc.

This NVR series, as an example, is a competitor to conventional analytics from the likes of Briefcam, Agent VI, ObjectVideo (RIP), Avigilon/VideoIQ, 3VR and the other 60 video analytic providers, most of them which still struggle to make profits and significant revenue from analytics.

Also, as for your 'sky high' margins, there is no doubt that the likes of NEC have sky high gross margins and profits but they have a ton of cost to cover as well, given these are high-end projects with highly paid business development / sales people plus engineering experts like yourself who work closely with customers, system optimization, etc. Hikvision's NVR here is a distribution box sale. Maybe one day deep learning is so good that it obviates the needs for professional sales people and experts but that's a lot to expect of Hikvision in the near future.

Avatar
Heng-Kuan Lee
Jul 21, 2017

Dahua had released "Deep Sense" server, it supports structure analysis of up to 192channels of video.

The GPU they used is nVidia Tesla P4, I think it's much higer performance than DeepInmind's GPU.

(1)
UM
Undisclosed Manufacturer #3
Jul 20, 2017

From a technical standpoint, it is child's play to use Deep Learning to filter out false motion alarms by rejecting images that don't include a person or vehicle for instance. I am surprised this is not offered by all NVR manufactures by now as it is easy to do, and you get to stamp the phrase "Deep learning" on the product which has some marketing value I guess. But it just scratches the surface of how deep learning could truly revolutionize the analytics side of the industry.

(7)
Avatar
Ricardo Souza
Jul 20, 2017
Motorola Solutions • IPVMU Certified

Eagerly waiting for some IPVM tests. Especially after those accuracy claims.

(2)
(1)
LS
Luis Sadio
Jul 31, 2017
IPVMU Certified

yes, with many different shadows and in a rainning day.

(1)
UM
Undisclosed Manufacturer #4
Jul 20, 2017

Hmm - reading the manual shows that you have to have special analytic cameras to make this work...but will wait for your review.

 

(4)
JH
John Honovich
Jul 21, 2017
IPVM

#4, good point. I missed that. I emailed Hikvision asking them what limits / cameras are supported.

(2)
(1)
U
Undisclosed #6
Jul 22, 2017

This sentence may need adjustment based on Hik's response:

"One structural advantage of them releasing a deep learning NVR is to make it easier to use existing / conventional cameras, rather than buying specialized ones."

UM
Undisclosed Manufacturer #7
Jul 23, 2017

I wonder if they have moved more into the Movidus solution.

https://techcrunch.com/2017/07/20/movidius-launches-a-79-deep-learning-usb-stick/

Hikvision was already into this solution:

https://www.youtube.com/watch?v=E5RkXGiKkDc

Good thing it can be an even playing field, with open SDKs/APIs can only mean lower cost.

So who's for "AI ready IPC" plug in and get the benefits.

Not so much shown but video below need just to 3:48 almost the end.

https://youtu.be/4xud1T9DaFY

Interest how classifiers can be used. I think license of many current technologies are very much way too expensive to people. This looks like it's paved the way forward.

$0.02

(1)
IK
Itamar Kerbel
Jul 23, 2017

IPVM is going to review/test this once this is out - right?

JH
John Honovich
Jul 23, 2017
IPVM

Yes, we mentioned that in the post - "IPVM plans to test the DeepInmind NVRs when released."

JH
John Honovich
Jul 24, 2017
IPVM

Updated Hikvision responds on Camera Support

Upon launch, the appliance with support streaming video from Hikvision plus 3rd party cameras via ONVIF. Initially, the appliance will only support human body detection from Hikvision Value and Smart Pro series. Future firmware upgrade will add deep learning support for 3rd party cameras.

U
Undisclosed #8
Jul 25, 2017

Any idea what value the NVR is bringing, if it relies on the camera to perform the analytics?

I'm assuming they aren't just slapping "deep learning" on their marketing materials while doing nothing more than forwarding edge-based analytics from the camera.

Is it simply trying to enhance the edge-based analytics by receiving alarms and doing further analysis to filter out false alarms?

Feels like bait and switch to say the NVR is doing deep learning/analytics for body detection but that it only supports the feature if the cameras themselves support body detection independently.

(1)
JH
John Honovich
Jul 25, 2017
IPVM

#8, sorry that is my fault. I did not make it clear where the different parts are being done. Hikvision described it as "The VCA function should be enabled in the camera. The Human Body detection is done in the NVR." I've updated the report below to clarify this.

U
Undisclosed #8
Jul 25, 2017

Okay, that's what I figured but wanted to be sure. Sounds like the NVR is using VCA metadata to enhance server side analytics which is a pretty common strategy.

Avatar
Mick Brown
Jul 25, 2017

Does this also mean hikvision can exploit the video from

their p2p servers further garner their deep learning

profiling westerns 

do we trust hikvision this

much

 

(3)
JH
John Honovich
Jul 30, 2017
IPVM

Update: Hikvision's DeepInmind brochure explains that the deep learning element will provide a false alarm filter to conventional alerts, as demonstrated in the screencap below:

Historically, a major problem of false alerts is that obviously (at least to human) false alerts would be triggered on leaves or cats or wind or shadows, etc. What Hikvision is claiming to do here is to take those alerts and see if the alerts contains any humans using deep learning. This could significantly improve performance.

(1)
Avatar
Heng-Kuan Lee
Aug 01, 2017

Hikvision co-works with Intel and nVidia to explore the possibilities of Deep Learning for the surveillance industry.

Hikvision together with Intel had released some storage technique recently.  It is very similar to DeepInmind brochure material pp. 5.

(1)
U
Undisclosed #10
Aug 01, 2017
JH
John Honovich
Aug 01, 2017
IPVM

#10 thanks for sharing that video. I embedded it to make it easier for others to see.

That noted, the video is poor evidence of Hikvision's face recognition quality. This is not a criticism of Hikvision's products but whomever did that 'test'. Take 3 people in an office, have them all look at a camera, and yes, face recognition 'works' but that's not realistic at all. You need to test it with lots of people, multiple cameras, multiple weeks, etc. to even begin getting a sense of a face recognition systems performance.

(1)
IK
Itamar Kerbel
Aug 01, 2017

Sorry but I'm missing something.

The first 3 people it detects fine? why it keeps changing the sample picture?

IK
Itamar Kerbel
Aug 01, 2017

Sorry - got it.

Its the same person from different years.

Still I don't think it should change the sample image...

Avatar
Heng-Kuan Lee
Aug 02, 2017

Is there any privacy issue in such product?

For general users, the NVR should be used to monitor current/recent status but not used to find the relationship of pictures including other information, e.g., young look.

Avatar
Mick Brown
Aug 01, 2017

More outlandish statements 

the compute power for this is way more than they can realistically deliver

16 cameras of simultaneous analytics

but this spec will be tweaked into a way they can

where its being more used in access control setting rather than video detection

picking up the companies 20 employees rather than the wider billions on the planet 

normal over spec under deliver