Ambarella on Computer Vision and US Hikua Ban

By John Honovich, Published Sep 10, 2018, 12:14pm EDT

Ambarella, a widely-used video surveillance component supplier, is betting on the rise of computer vision and is already seeing a sales impact from the US government ban of Dahua and Hikvision, according to a recent investor's call.

cvflow us gov ban hikvision dahua

In this note, we share details from that, examining Ambarella's new chips, how Ambarella is hoping customers will use it to improve video analytics and who is winning and losing from the US government ban.

Ban ****** - ***** ***** ****

***** *** ********* **** been **** ** *** ***, says *********** * *** ******:

*.*. **** - **** S. ****** ******** ******* Authorization *** *** ****** Year ****, ***** *** recently ******* **** *** and **** ** **** effect ** ******** **** 2019, **** ********** ****** our *** ******* ***** IP ******** ********* *** likely ******** ***** ****** for *** *********.

*** ****** ** ******* happening, **** ********* *** Dahua ******* ****** *** 2019 ********, ********* **** in *** ********'* **** [link ** ****** *********]:

** * ****** ** the *.*. *** ** purchases ** ******** ******* from ********* *** *****, we *** ****** **** near-term ********* ** *** orders, ************ ** *** high *** ** *** product ***** ******** ********** with *** ****** ********.

*** ****** ******** *** Dahua *** ********* ** the ************* ****** ******** 'overseas' **** ** ** the ***.

*******, ********* **** ***-******* customers *** **********:

** **** **** ** offsetting ******** ** ****** from ***-******* *********, *** we **** *** *** the ******** ****** ** revenues ***** ***** ** the ****.

**** ****** *** ******* for ********** ****** ******* growth ** ************* **** project ***** * ****** to * **** *** to ****** *** ********* components ** ***** ******* in *** **** ******.

4K / ****** ********** ********

*********** ************ ** ****, 4K*** ****** ********** ****** have **** *************. *************** ***** ********* ** increase ********, ** ****** ** higher ********** * ******* growth ******.

Computer ****** / ** / *****

***, *********, **** **** in *** ******** *** shifted ***** ** ******** vision, ** **********,  ***** ****** *** **** chip, **** ********* ******:

** ******* **** *** security ******* **** ******** vision ********** **** ******** to **** ** * percentage ** ***** ************ IP ****** *******, ********* those *** *** ************* as **** ** *** replacement ** ******** *******. This **** ******* *** opportunity *** ** ** increase *** ***** ************ per ****** *** ** grow *** ***** *******.

*** ********* ******* *** users ***** ** ** run **** ******** ** the **** ***** **** do ***** ********, ******** the ****** ** *** only ****** ************ ***** but ** ******** *** categorize ******, ****, *******, e.g., ** ********* *** false ****** **** ******, shadows, *********, *********, ***. that ****** ** **** security ******* (****** ********* ******** *******). 

*** **** **** **** is ****** ** **** a ****** ** **** announcements ** *********'* ********** on ****** ************ **** orders ** ********:

** *** ********* ** the ******** ******** ** our *** ******** ****** SoC ** ** ******** camera ****** *** **** forward ** ********* ***** in **** ********** ** the ***** **** ** next ****.

*******, ***** ******* *** going ** ** *********, as ********* ** ********** a * ** ** increase ** **** ***** component *******:

******** ** *** **** starting ******** ** *** price ************ **** *** customers, *** ******* ** CV22 **** ** ** to ** ****** **** our ******* ******* **** in *** ******** ****** market, *** *** ** probably **** ****** **** that.

***** ****, ********* ********* that, *** ***, **** will ** *** ****-***, expensive ******* *** **** have ** ******** *** chip ******* ******** ***** costs ** *** ******:

* ** ***** **** you're ********** ** *** category ******** *** **% of *** ****** ***** now ***** ** *** current ****** ***** ** security ****** ****** *****. But *** *** ********** right ** **** ** are ********* *** *********** price ***** ** ****** out ******* **-********* ****, because ** ** ****** that *** ** ** continue ** ****** *** transition **** *** ***** to ** ********, *** customer **** ****** ***** price, *** *********** ******** that ** **** ** be ******** *** ****.

** ******, **** ** Ambarella ******** *** ******** to ******* ****, *** manufacturers **** ** ******* and ******** *** *********** models ** **** ********* of ****. 

**** *********** *** ******** computer ****** ********* *****, with************** ********, *** *** **** notable. *********'* *********, ********** else *****, ** **** **** large ****** ************* (**** Avigilon ** **** ** Bosch ** *********, ***.) already *** ***** ********, making ** ****** ** continue / ****** **** them.

** *** ***** ****, Ambarella **** **** **** Dahua *** ********* *** working ** ***** *** chips, ******, ** ** not ***** **, **** or *** **** ***** chips ***** ****:

* **** **** ********* and ***** *** ** have ******* ** ***** chip, *** ******* *** years, ** ********* ** be ***** ********. **, I ***’* **** *** the *********** **** *** customer ********* ** **** on **, ****’* ********* we ******** ** **** with, *** * ***** when *** ** **** comes ***, **’** ****** raise *** *** ** this ****.

Growth / ******* *********

*********'* ***** *********** ** ****, ******* by **** **** */*** since ****, ** ***** previous *** ******** ***** hit **** ***** *** shifted **** **** *********.

***, *********'* **** ********* market ** ******** ( i.e., ***** ************) ** the ******* ***** **** than **** ***** **** our ********:

******** ** ***** ** going ** ** ** least **% ** *******. And ****, **** **** be ** *** ***-***, so **** ****** ******** and * ****** *** of ************** *** *** remainder.

*******, ******* * ******* of *** *********'* ******* comes **** ***** *** Hikvision ** *** ******* notes ** *** **%:

**’* ******** ***** **% into ***** *********, **% into ***** **** ***-******* customers.

[****: ** ******* ******* estimated * ***** ** Ambarella's ******* **** ***** but ** ********** *** not ****** *** ******** vs ************ ***** ** the **% ******** *****.]

****** ** *****, ********* is ********** ** ******* with *** *** ***** costing ****** ********* ****.

*******, ****** ** ********, consumer ******** (**** **** or ****) *** ******** stronger **** ** *** professional **** **** *** no ******:

** ** ********, ****-****-**** growth *** *** ** consumer ******** ********, ***** professional ******** ******** **** flat.

***** **** ****, ********* will **** **** ** navigate *** ***'* ******* and *** *** ********* / ******* ***** ********* are ** ********** ******** vision ***** ******* **** work *** ***** *********** demand.

Comments (13)

Nice reporting.

Any idea what % of Hikua’s cameras are Ambarella powered?  

Related, are there any other major options besides HiSilicon, Sony and TI out there?

Would it be possible during the IPVM product review process to inspect and note the imager/SOC chipsets in use for all products?  This information is not readily available anywhere, and could become more important to buyers, as cyber-security concerns will demand further transparency into what is actually in the box.

Allwinner
Socionext
Rockchip

Hanwha uses their own Wisenet Chip
Arecont uses FPGA of their own design

I'm sure some others can identify a few more.

#1, Ambarella lists the companies in their financials they see as their main competitors:

In the IP security camera market, our primary competitors include Fullhan Microelectronics Co., Ltd., Geo Semiconductor, Inc., Grain Media, Inc., which was recently acquired by Novatek Microelectronics Corp., or Novatek, HiSilicon, Intel Corporation, or Intel, Movidius Ltd., which was recently acquired by Intel, OmniVision Technologies, Inc., or OmniVision, Qualcomm Incorporated, or Qualcomm, Realtek Semiconductor Corp., Socionext, and Texas Instruments Incorporated, or Texas Instruments, as well as vertically integrated divisions of IP Security camera device OEMs, including Axis Communications AB and Sony.

Fullhan is notable since Hikvision's initial investor and Board member Gong Hongjia is also a director at Fullhan.

Last year, Hikvision spent ~$32 million USD double what they did in 2016, references:

As for:

Any idea what % of Hikua’s cameras are Ambarella powered?

Good question. I don't know but I did some looking around. So last year, Hikvision said:

In 2016 Hikvision manufactured more than 55M cameras, while in 2017 H1 the company produced in excess of 30M camera units.

Let's say they made 70 - 75 million cameras last year, assuming growth in H2, especially since H1 includes Chinese New Year.

Hikvision 2017 financials says front-end equipment (i.e., cameras) generated 21,090,230,299.49 in sales in 2017, which is ~$3.1 billion.

Combined that means the average sales price per camera is ~$43 (i.e., 3.1 billion revenue / 72.5 million cameras).

~$43 per camera is quite low, certainly by Western standards but could be accurate depending on how low they sell to OEMs, OECD world distributors, etc.

However, if that is anywhere near true, this would imply they do not use a lot of Ambarella SoCs as they are at the high of the market and would be hard to fit when selling cameras at sub $50 level.

Anyone see any issues with math or assumptions, let me know. Curious to hear others thoughts here.

I hope Ambarella is not betting too much on the security industry doing more with edge analytics any time soon. While their CV-optimized chips may make part of the equation a little easier, I think we are still quite far away from edge analytics being "common" in surveillance cameras.

One significant issue holding things back is the lack of uniformity on the VMS side. Avigilon has good analytics integration (setup, alarm handling, searching, etc.), but they are really closed to supporting outside products in any way as fluid and simple as their own cameras. For all the other VMSes analytics support varies from practically non-existent to mostly alarm-receipt-only. 

Until we see VMSes more widely supporting setup of rules ad events, searching meta-data, displaying dynamic meta-data and other things that tend to be more specific to analytics cameras, I don't think we are going to see major growth in this segment to the extent it bails out Ambarella (or any other chipmaker).

Also, it is not like there haven't been other CV-enabled chips before now, so this announcement from Amabarella is not really a market disrupter.

Ambarella, and some of the other chip companies talk about how "easy" it is to port analytics to their chips using open-source tools and such. If they really want to drive adoption of these chips they should offer subsidized engineering services to help companies port old algorithms to these new chips. There are a LOT of little analytics companies out there, many would likely welcome help from Ambarella porting over to these chips. They might even be able to setup some kind of licensing arrangement where customers of Ambarella could get access to these analytics for a small incremental price, that way a camera company that doesn't have their own analytics could build them into their designs more easily.

This is an important point:  edge analytics will struggle until VMS's agree on how to handle the data.  ONVIF is way behind in this effort. 

Does this mean that video analytics will be relegated to the NVR/DVRs?  If so, how is the decode load going to be handled?

On a different note, while it's true that many video processors have offered Computer Vision, the benefits of Deep Neural Networks appear to be real.  They can reliably find and track objects and distinguish between 'car', 'person', 'leaf', 'squirrel' and 'raindrop'.

 

On a different note, while it's true that many video processors have offered Computer Vision, the benefits of Deep Neural Networks appear to be real. They can reliably find and track objects and distinguish between 'car', 'person', 'leaf', 'squirrel' and 'raindrop'.

DNN's are real, but they don't necessarily require specialized processors. There have been products with edge-based DNN analytics running on standard SoC's and DSPs...

Also, the greater the variety in object types you want to classify, the more power/compute demand increases and the more chances of false alarms and/or missed classifications increase. A system built to distinguish between car/person/leaf/squirrel/raindrop in the state of technology today would likely not be a low power/low cost edge-based solution.

Clicked Agree but would like to make a case for where I disagree with your very last sentence. 

My premise is that in a majority (not all) of security/VCA applications, the goal is to accurately classify person(s) and/or vehicle(s) and run VCA rules against these objects to create accurate alerts/notifications and minimize false alarms. 

Classifying everything else from squirrels in the perimeter to spiders on the cameras to swaying bushes/branches as a say, <null> object, clutter or activity to be ignored could help greatly in the reduction of false alarms without burdening the processing overhead.  Long story short, one would not have to create separate deep learning classifiers for common causes of false alarms, thereby reducing the processing burden on the edge.

Classifying everything else from squirrels in the perimeter to spiders on the cameras to swaying bushes/branches as a say, <null> object, clutter or activity to be ignored could help greatly in the reduction of false alarms without burdening the processing overhead. Long story short, one would not have to create separate deep learning classifiers for common causes of false alarms, thereby reducing the processing burden on the edge.

But <null> object is essentially the default classification. The high level approach is that any object classification analytics system is using a series of filters to go from "moving blob of pixels" to "classified object".  A first-pass filter is usually size/shape, the system uses calibration data (either entered manually, or determined automatically) to decide when a blob of pixels is either too large, or too small, to be a potential object of interest. From there you often look for things like movement/direction, texture map, aspect ratio, size relative to location in FOV, etc. The actual "deep learning analysis" part is applied after the blob has already gone through the lower-cost (from a resource perspective) high/low pass filters. Anything that does not make it through those filters is therefore a null classification by default.

Attempting to classify more non-human objects would not necessarily reduce false alarms. It would just increase processing load on the system. Keep in mind that many times an analytics classifier, whether edge-based or server-based, is architected based on assumptions that only X moving blobs need to be fully classified at any given time. If you try to classify too many things, you are either going to need to up your hardware resources, or risk missing a valid object while you were spending too many cycles concentrating on the squirrel.

 

I agree with your first paragraph.  This is a very good description of the general approach taken by VCA vendors to classify objects and how most will likely choose to implement deep learning/'X' neural networks to improve classification.

I'm a bit puzzled by your second paragraph as it seems to bring us back to the original reason I had an issue with your previous comment's last sentence: "A system built to distinguish between car/person/leaf/squirrel/raindrop in the state of technology today would likely not be a low power/low cost edge-based solution."

You finished your reply with "If you try to classify too many things, you are either going to need to up your hardware resources, or risk missing a valid object while you were spending too many cycles concentrating on the squirrel."  True, but what I'm saying is screw the squirrel!  It's not a person, it's not a car, I don't care, it's noise and I don't have to spend my cycles trying to classify and track it while it enters and exits my no parking zone or perimeter line.

I appreciate we may be talking past each other due to terminology, some fundamental misunderstanding or different engineering approaches our companies may take in implementing deep learning in future products.  However, I will take a gentleman's bet of a beer at the next ISC that the IPVM team will have several interesting deep learning enabled, non-GPU dependent appliances (i.e.not high end, high price servers) and cameras from multiple vendors within the next two quarters with improved classification, tracking and false alarm performance than their recent video analytics test.  Whether they fit your low-cost criteria for edge-based solution is something we'll have to quantify before the next round.

cameras from multiple vendors within the next two quarters with improved classification, tracking and false alarm performance than their recent video analytics test.

You recognize that for the above to be true the products would have needed to have begun development at least 1 year ago from now, right?

News on CV adoption: Avigilon has announced their upcoming AI H5 line. However, though Avigilon is an Ambarella customer, Avigilon has chosen Intel Movidius for this.

update: as you point out here, and in support of what you posted below, Avigilon ultimately chose Amberella over Intel.

Update: On the Computer Vision side, Ambarella says 5 of their professional security camera customers to ship in 2020:

That sounds reasonable to us since those chips have been available for about a year, giving the typical time to do development / integration.

Ambarella also listed the following customers:

Read this IPVM report for free.

This article is part of IPVM's 6,735 reports, 909 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

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