Avigilon Appearance Search Tested

By Rob Kilpatrick, Published Oct 30, 2019, 10:32am EDT (Research)

******** ********** ************ **** ** "***** ******* ***** of ***** **** ****, ** ******* locate * ******** ****** ** ******* of ******** ****** ** ****** ****, or ******** *****." *** **** ** really **** **** ******** ** ******* search?

free image2 2

** **** ***, ** ****** ***** claims *** ******, *********:

  • ** ** ****** **** ********* ******* footage ** * ****** *** ***** timeline ** ********* ******?
  • ** ********* ** ******** ****** ********?
  • *** ******** ** ** ** **** light, *** *****, *** **?
  • **** ** ********** ******** *** *** gender?
  • *** ******** ********** ********** *** ********?

Executive *******

***** ** *** *****, ********** ****** provides * ************* ****** ***** ** search *** * ******** ****** ** vehicle ****** ******** ******* ******** ** using ******** ** ********* ****** *****. However, ***********/******** *********** **** ** ***, gender, *** **** ***** **** ******* inaccurate ** *** ***** *** ********** IR (*** ********* ******) ****** ************* ** ******* *** appear *****/****.

Key ********

** *** *****, ********** ****** *** several *********:

  • ****** **** ******* *** ******:***** ********** ****** *** ************* ****** than ********* ******* ********* ** ***** typical ********* ******, ********** ****** ******** cameras *** ******** ****.
  • ******** ***** ****** *******:********* *** ***** ** ***** **** color *** ********* ******** ** *** tests ** ******* *** ******** *** light ******.
  • ******** ******* **************:********** ****** ********** ********** *** ******** multiple ******* *****, ********* ****, ******, buses, *********** *** ********.
  • ****** *************:** *** ********** ******, ***** ****** check ** *** ** ****** ** on **** ******. ** ******* ************* is ********.
  • ** ********** **** **** ***:********** ****** ** ******** **** ******** Control ****** ********** ** ** ********** cost. ** ******* ********* ** ********.

*******, ***** **** ******* *********, ** well:

  • ********** *********** ***********:********* ** ****** ** *** (*****/*****) were **********, **** **** ******** ***** as ****** *** **** ***** *** adults ***** ** ******** *** **** versa.
  • **** ***** ** **** **** ******:**** ********* ***** ** ***** ****, subjects **** **** **** ********** ***** in ****** *******.
  • ********** ** *** *******:** *** **** **** ** **, all ******* *********** ****** ***** ** gray, ******** *** ****** **** *** be ********. ******* ** ******** ******** in ** **** **** **** ******** than ***** ** ***** ******.
  • ********** **** *******:** *** *****, ********** ****** ********* detected ********* ******* (***** ** *****) or **** **** ** * *******'* face **** ****** ** *** "**** Profile." ***** ******* ****** ** ******** adjusted ** ***********.
  • ******** ********* ****:********** ****** ***** **** **** ********'* own *********, *** ***** *******, ************/********** ** *** ** ***** ********* appliances (* *********** ********** ****).

** *********** **** ** ***** **** and **** ** ********** ****** ** this *****:

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

** *** *****, ********* ***** *** lower **** ******** ****** *** ******** during *** *** *** ** ******** low ***** (~*-* ***). ******** ****** were **** ******** (***, ******, ******, blue) *** *****, ****, ***** **** accurate, ** ****.

Search For Upper Lower Body Color Accurate

Subject ****** ********* ********

********* * ****** ** * ******** subject **** ******** ***** *** ********* accurate ** *** *****, **** **** results ******* *** **** *******. *** example, ******* ***** **** *** ******* on *** ********* ******* ** ***** different *****.

Specific Person Search From Recorded Video

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

********** ****** ******** **** ********* ** one ********* ** ****** ******. ** a ****** *** ******** * **** considered ******, ** ** ***** ** the *******'* ********** ***********, ***** *****. In *** *****, **** * **** was ********, ******** *** ********* ******, especially ****** ******** **** *** ******** changes. *** *******, *** ******* ** wearing ***** ********* ******/**** ** **** image:

*******, **** *** ** ********* *** users ** ***** ** ** *** to ******** ****** ***** ***** ***/*** not *****, ***** ** ** ********** of ******* * **** ** ** is *** ********, *** ***** ****** be ***** *** *****. **** **** is******** ** ** ********** ********* ** the *******'* **********, ******* ** ******** color, ******/***** *****, ***.

************, ***** ** ** *** ** mark * **** ** "*****" *** tell ********** ****** ** ******* **. This *** ** ** ***** ** the ******** ************ ******** ********* ******* such ** *** ***** (**** *****) or *****, ***** ****** ** ******* from *** ******* *******.

Floor Detected As Face  No Way To Mark Incorrect

Searching ***/****** **********

**** ********* ***** ********, ********* ** age *** ********** **********, ******* ****** adults **** ********* *** ********.

Searching For Children Displays All Adult Subjects

************, ********* ** ****** *** **** inaccurate, ********* ****** **** ******* **** high ********* **** ***** **** ********.

Searching For Female Displays Several Men

Searching **** ***** **********

********* **** ***** ******** ***** *******, for ******* ********* ******* **** ***** hair ***** ******* ******* ** ****** wearing ***** ****, ********* *** ****** people ***** **** ******.

Search For Dark Hair Displays Subjects In Hats

Accuracy ******* ** ** ******

********* *** ******** ** **** ****** with ********** ** ** ******** **** accurate ******* ** *** *******'* *****(*) can *** ** **********. *** *******, despite ***** ****** ** *** ******* in ** *****, **** ********** ******* of ****** *** *****.

Accurate ******* ****** *** ***** / ****

** ******** ** ******* ********, ******* color ****** **** ****** ****, **** red **** ***** ***** **** *** course ** ******* ****.

Vehicle Color Accurate

******* **** *** **** ********** *********, with ****** *** ***** ************** **** cars:

Trucks Accurately Classified Searched

*** *********** ********* ********** ************** **** bicycles ** ***** ********:

Motorcycles Accurately Classified Searched

Simple *****

** *** ** ********** ******, ********'* Analytics ******* **** ** ********** *** installed ** **** ****** (~* ****** process). ***** ************, ********** ****** *** be ******* ** ******* *******.

Pricing / *************

******** ********** ****** ** ******** ** ACC ********** ** ** ********** ****. Users **** **** ******** *** **********'* ******** ******** *** ** * ***** party **** ****** **** ********* *** (Nvidia ******* ** ******).

********** ****** ***** **** **** ******** Analytics, ****** ******* (*** *** ****)**** *********** ** ************ ***** ***** *******.

********

******** ******* ****** *.*.*.** *** **** in *******. **** ********* ** *** H4 ******** ******* **** **** ** testing, ********* ******, ****, ***, *** multi ******.

Comments (30)

Regarding the IR search problems, this is a general issue with IR and color based analytics, of course, but I think it could have a general impact on camera selection. The more people make use of such analytics, the more there will be a push for non-IR / low light cameras.

Agree: 7
Disagree
Informative: 1
Unhelpful
Funny

We have Appearance Search deployed on several large systems in many different scenarios with hundreds of cameras each and I am continuously impressed on how well it works.

Agree: 4
Disagree
Informative: 4
Unhelpful
Funny

Michael, any specific examples? That's a pretty abstract endorsement and if you want to promote your partner that's ok if you back it up with specifics.

Agree: 4
Disagree
Informative
Unhelpful
Funny: 3

Ok, one specific example we just installed a Vape detector in a school which is tied to an Avigilon H4A camera outside the bathroom. I just did a search after a vape alarm came in and I was able to right-click and on the person leaving the bathroom and Apperach search showed me when they entered the bathroom and which classroom they came from. In a school with 300 cameras, it was impressive how fast I was able to find where they came from.

Agree
Disagree
Informative: 17
Unhelpful
Funny

is the kid “tagged” with the vape alarm?

Agree
Disagree
Informative
Unhelpful
Funny

No... A) alarms do not get locked to people/vehicles as a human needs to select them in the software B) there are no cameras in the bathrooms so the camera doesn't see who triggered the alarm only people outside the bathroom.

Agree
Disagree
Informative
Unhelpful
Funny

so in this particular case, if the same sequence of events occurs tomorrow, finding that same person wouldn’t indicate the alarm from the day before?

Agree
Disagree
Informative
Unhelpful
Funny

Alarms have nothing to do with Appearance Search besides seeing the people detected in the alarm event. So if I do an Appearance Search over a 48 hour time frame the system will look for that person or vehicle over the whole 48 hours regardless of how many alarm events there are.

Agree
Disagree
Informative: 2
Unhelpful
Funny

I know when we tested AnyVision (Different product, but similar, I guess?) they claimed the system would only get better over time. As it sees more, and learns, especially with user input (yes this face is actually that person, no that is not a face, etc).

Do you think this is potentially true of Avigilon as well? I have some doubts as you mention you cannot tell the system that the picture of a puddle is not a face, or that his face is not the same person as the other one.

Does Avigilon plan on fixing that?

Agree
Disagree
Informative
Unhelpful
Funny

Avigilon is different then Anyvision in the fact that once you do an Appearance Search you can select the results that match your search and star them. Every time you star a result the system searches again and the more stars you have of the person or vehicle you're looking for the better the results are.

So you will get false positives in the initial search but the more you star your subject the better the results get. I don't think this was explained very well in this test.

Agree
Disagree
Informative: 4
Unhelpful
Funny

Additionally, AnyVision is true facial recognition. Avigilon is classifying humans and vehicles. Appearance search is looking for matching patterns of classified objects rather than facial recognition. They're very different.

Agree: 7
Disagree
Informative: 1
Unhelpful
Funny

Just curious, what's a passing grade to earn a green checkmark vs. red 'x' on these tests?

For example, the motorcycle classification is called "generally accurately", which is fair, but I wonder if giving the green check on ~70% accuracy(7/24 images shown are clearly not motorcycles) is too generous. Of course you're probably looking at a larger data set in your eval, but it's tough to get that out of the report.

Agree
Disagree
Informative
Unhelpful
Funny

Thank you, great review.

It would be great to see some measurable test of Avigilon Self Learning, for example based on Gender or Age.

Also would be great to see how many results were missed vs found. For example Truck Classification - how many were found vs missed.

Agree: 2
Disagree
Informative
Unhelpful
Funny

Again keep in mind with Avigilon your initial search results are not going to be 100% perfect. You as the human star the results and the system then looks for those results. The more you star a person or vehicle the more accurate your search will be.

Agree: 1
Disagree
Informative: 4
Unhelpful
Funny

I am a huge fan of Avigilon and we use Avigilon appearance search extensively. However, while is has brought about some efficiencies when searching for simple things, it is definitely not accurate on gender or age and the results are even worse when the subjects are of a different race. Not only has it been consistently inaccurate in the search results, there are many misses. Self learning has absolutely not improved that over time.

Agree
Disagree
Informative: 4
Unhelpful
Funny

Are you talking about the initial search results or after you star your results?

Agree
Disagree
Informative
Unhelpful
Funny

Both initially and after results are starred.

Agree
Disagree
Informative: 1
Unhelpful
Funny

I can't say I agree with you when it comes to the starred results. Also, self-learning is for the analytics in the camera, not the Appearance search.

Agree
Disagree
Informative
Unhelpful
Funny

Yes Michael, it's very interesting.. That's what i want to see as a measurable statement (test results).

Agree: 2
Disagree
Informative
Unhelpful
Funny

I suspect this is not measurable in the scientific sense of hypothesize then verify through experiment. There are too many variables which can impact the probability of accuracy.

In any real world scene, if you move the camera left or right or up or down the results of an experiment could be vastly different (or not) when it comes to analytics.

It seems to me the essence of video analytics is increasing the probability of detection of undesirable circumstances in the field of view of a camera in order to be more proactive in real time assessment and, in the case of appearance search, massively reducing investigation time even with the certainty of false positives. It's much easier to investigate something that might be wrong rather than investigate something you never would have known about without tireless hours of scrubbing video.

Agree: 3
Disagree
Informative: 1
Unhelpful
Funny

I would disagree with your first statement in sense of probabilistic model being immeasurable, but that's not the point here.

It's much easier to investigate something that might be wrong rather than investigate something you never would have known about without tireless hours of scrubbing video.

Yes, exactly. Main point would be to filter out events based on your criteria, but not to miss events. Let's assume Truck is the criteria in this case, and we only want to get Trucks. Test wise - I would like to know how many trucks we missed vs how many trucks we got correctly, does self learning improves this ratio. Does thumb up/down event helps or overfits model and makes it worse in general case?

From customer standpoint: It's easy to see correct/incorrect events in result set presented in UI. Its impossible to know based on this result set what is missing from it. How would you know you missed the Truck you are actually should looking for, would you scroll through hours of video?

Agree
Disagree
Informative
Unhelpful
Funny

This is what I see as the future of camera systems and they’re ahead of the competition.

It’s a pity that it only works with their cameras and it’s expensive.

Agree
Disagree
Informative
Unhelpful
Funny: 2

FYI you don't have to use Avigilon cameras for Apperence Search. Also I disagree on the expensive part as Appearance Search is no extra charge unlike other systems that require additional licneses and servers which adds significant costs compared to Avigilon.

Agree: 2
Disagree: 1
Informative: 2
Unhelpful
Funny

Expensive is relative. A significant part of the market has been accustomed to $50 - $100 cameras, making this expensive for them. I am not arguing against the value of this, just pointing out that from a Hikua customer perspective, Avigilon looks very expensive. And I don't know about #4, but I am sure some integrators will be reasonably hoping "Hey Hikua will have this in {insert months} and will give it away for free."

Agree: 2
Disagree
Informative
Unhelpful
Funny

Totally agree but currently to get similar features you are bolting on additional software which requires more servers and licneses which adds significant costs compared to Avigilon.

Agree: 2
Disagree: 2
Informative
Unhelpful
Funny

Sure, compared to something like Briefcam. That's why I think it's relative. If you are coming from the "I can spend lots of money for a premium solution" position, Avigilon Appearance Search may seem inexpensive but if you are coming from "I don't pay a penny more than $100 for cameras" this will.

Agree
Disagree
Informative
Unhelpful
Funny

Avigilon did well acquiring a good analytics company and getting a jump start. In the coming months, there will be other reputable manufacturers (at least 3) with machine learning products that are open, edge-based, and work extremely well. A few months after that release, major VMS companies will accept and integrate the metadata from these cameras and object\color\feature search will be mainstream. I doubt Avigilon will comply to these camera manufacturers, at least until there is surmounting financial pressure ( like adopting H.264 and ONVIF).

I wonder how Avigilon will stand up to the combined sales force leveraged through a vast channel (distribution, manufacturer reps, integrators, e-commerce, tech partners) that they don't have (unless you count Motorola's radio channel?) when their edge disappears.

In 2020 when these open edge AI products hit the market, I'm not sure how these AI software developers will grow (Anyvision, Breifcam, etc) when the base license and\or recurring fee is so high per camera and restricted to using dedicated server hardware.

The desire to eliminate servers is a very strong trend among end-users currently.

Maybe Hikau will become unbanned and unsanctioned and take over with AI, who knows??

Agree
Disagree: 1
Informative
Unhelpful
Funny: 1

In the coming months, there will be other reputable manufacturers (at least 3) with machine learning products that are open, edge-based, and work extremely well. A few months after that release, major VMS companies will accept and integrate the metadata from these cameras and object\color\feature search will be mainstream.

You want to bet a $1 that it is going to take way longer for this to happen? And to be clear, 'machine learning' is pretty broad, if you mean person detection, sure I agree, but we in this post, we are talking appearance/similarity search. That I think is going to take a lot longer to be 'mainstream'.

Agree: 2
Disagree
Informative
Unhelpful
Funny

It can also sort of be hard to get Avigilon products. Maybe not as an end user, but for example I've wanted to start trying to maybe sell some Avigilon with the side jobs I do now and then. I can't call Avigilon, as I'm not large volume enough to be a dealer, the other local dealer will possibly sell to me, at a massive increase over MSRP. Either that or they deem me low priority and never get back to me. I've run into this with more than just Avigilon. If I'm not spending big money, no one(dealers) seems to be in a hurry to get back to me with pricing.

So not only is Avigilon already a more expensive product, I get stuck paying even more since I can't become a dealer, or waiting for ever.

Agree
Disagree
Informative: 3
Unhelpful
Funny

I firmly believe everyone must start somewhere. Some of our best dealers came from beginnings just as you describe. I encourage you to reach out via partner.program@avigilon.com and request to become a dealer if you are still interested.

I look forward to speaking with you about the challenges in the past and what we can do to change the future.

Agree: 2
Disagree
Informative: 3
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