Axxon Face Search Tested

Avatar
Rob Kilpatrick
•Published Oct 26, 2018 14:12 PM

************ ******* ****** *********** ** ********** **** ****** **** **** *** ****** ********* ******, ******** ** ***** ***** ** find ******** ****** ******** ******* ** drastically *** ****** ****.

ipvm tested 2

** **** ******, ** ******* *** full **** ******* ** *****'* **** Search, *********:

  • *** **** ***** ********* ** ********* to ******?
  • *** **** **** *** ******** ****** subjects ******* **** ** **** **** the ******?
  • *** **** *********** ******* ** *** light, <1 *** **** **, *** *** scenes?
  • *** **** **** ********* ****** *** usage?

******* ********** ** ****** *********** ****** also ****** ****** **** ** *****'* **** Recognition ******, *** ********-**.

Summary: ******** ****** ***** ******* ** ****** ******** 

***** ** *** *****, ***** ****'* Face ****** ****** ****** **** ** default **% ********** ********, **** **** subjects ******** *******, **** *** ****** face ********** **** * **** ** testing ** * ***** ****** ** ~30-40 ******. ***** **** ****** *********** cuts **** **** ******** ** **** specific ******** ****** ******** *******.

************, ********* ********** ****** ** ***** light *** *** ********** **** ***** face *** ************ ** **** ***** (see *** **** *******'* **** *** ******), **** ***** ***** ****** ********** in *

Some ******* **** *** *********

*******, ***** *** ******* *** ****** to ** ***** **:

  • **** ***************:**** ******** **** ********** ******** *** others ** *****'* **** ******, ** high (>**%) ********** *******. **** **** often ******** **** ******* *** ** our *****. ***** ***** ********** ********** (~60-70%) ******** **** **********, ** ****, but ***** ****** *** ***** *****'* default (**%).
  • ******* ***** ****** **********:** *** *****, **** ***** **** rec *********, **** ****** ********** ********* by ~**% ** **** **** ******** looked **** ** **** **** *** camera, ************ ******* ******* ****** ******** (~5-10% ** *** ****).
  • ***/*** ***** ********** **********:************, ********** *** ******* ** ~**-**% in *** ***** (~*.* *** - 2 ***), ***, ** **** ***** integrated **, ******* ** ***** ********* in **** *****.

Upcoming **** **** *** *********

***** ***** ****'* ****** *********** ** limited ** ******** *** ** ******* versions, ********* *** ***** ** ******* live ********* ** *** **** *** months. **** ****** ** ***** ******* of ***** ********* *** ***** *********** fairly ******* ** **** ******, *** it ** ******* **** *******/************ **** be ** ***** ********** ********. ** plan ** ****** **** ******** *********.

*******/************

**** ****** ** ******** ** ** additional **** **** ***** **** ******** licenses (~$*** ****/*******) *** ** ** add-on ******* **** ***** ***** **** Professional. ***** **** ***** **** *** support **** ******. ****** **** **** recognition ********* **** ******* ********** *********.

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

**** ****** ** ****** ** ***, with ***** ********* *** ** **** cameras, ********* * ********* ***** ** the *******, *** ********* ********** ********* (if *********). ********** ******** ** **%, which ** ***** ********** ******** ** our *****, **** *** ******** ************ of ******** ***** **** *-* ********, discussed *****.

*** ***** ***** ************ *** ** use *** ***** **** ****** *******.

Configuration ********

**** ****** ******** **** ********* ** be ******* *** **** ******, ***** essentially **** ******** *** **** ** be *****, ******** ********** *** **** width/height ******** ** *********, *** ********* enabled.

**** **** ******** ********** *** ******* impact *** ***** ** ******* ***** 1080p. *-*** ****** ** *** ***** caused *********** *** ******, ********* ******* below ** **** ******.

*** ********* ***** ***** *** ******* configuration ** **** ********* ** ** individual ******.

Solid ****** *********** ***** ******* ** ****** **********

***** *****'* ******* **% ********** ** search, **** ******** **** ******** **********. For *******, *** ******* ***** *** consistently ********** **** >**% ********** ****** multiple ******* ** ******** ********** *** at ******** ******, **** ** ***************** at ***** ********.

most subjects recognized with high similarity

Bearded *** *************

*******, ******* ******** **** ********** ********** as **** ***** ** * **** similarity *****. *** *******, *** ******* below ** ********** *** **** ** himself ** **% **********, *** ** another ******* *** ** *** **** level. ******* ******* **** **** **** other ******* *** ****** *******. **** that ** ***** ****** ******** **** as ******* ** **** ****** ***************** in *** *****.

bearded subject

Lower ********** **** ***** ******

*** ******* ****** ** **** ****** similarity *** **** ****. *** *******, when ******** ****** **** ** ** the **** (** **** ** **** walking ********), ********** ******* ** ~**-**% or ****. *******, ****** *** ****** test ** *****'* **** *** ******, this ****** ****** ******** ****** ****** detections, ********* week ** *******.

head titled 2

<1 *** ********* ********, *** ******* ******* **********

***** ******** *** ********** ***** ** ~0.50 ***, ******** ********** *** ******* by ~**-**%. **** ** **** ** contrast ** *** ***** ******, ***** failed ** ******/********* ******** ** lux.

low light 2

CPU **** ******

*****'* **** ********* ********* ********* **** by ~*-*% *** ***** ****** ***** analyzed ** *** *****, ***** ** Intel ** ***** *********. **** ******** in ~**% ***** *** ***** *** four ***** *******, * *********** ******** over ******* *** ******* **** *********, which ******** **** *-*% ****.

*******, *** ***** ****** ******* **** streams ***** ***** **** ********. ****** two **** ***** ** **** *******, 3 *** * **, ********* *** usage ** **-**%****, ********* ** **%+ *** ***** total. ***** ***** ****** ********** ******* or ***** **** **** ****** ***** should ****** ** ***** *********. ***** resolution ******* *** ** ********, ****** with ********* ***** *************** *********.

CPU ****/*** ******* *******

*********, ***** ********* **** ********* *** recognition **** ** *** ******'* ***. However, **** ***** **** **** **** to ********* *****'* ******** ******** ****** toolkit ** *** **** ****** ** increase **********, **** ***** ******* *** GPUs.

Wide ***** ** **** / *** *** *** **********

**** **** ***** **** ****** *** be *** ** *** ****** **** Face ********* *******, ***** **** **** fields ** **** ****** *** *** provide *** ********* ******* ** *** tests, *.*., ** *** ***** ** view *****.

wide field of view

***** ********* ************ ** ** ****** between *** **** *** *********, ** roughly ***-*** ***. ************, **** ** PPF ****** *** **** ****** *** analysis, ***** ****** ****** ** *** CPU ****** ** ***** ***** *******, as ***** ****** ****** ****** *** cause *** ***** ******.

Versions ****

*** ********* ******** **** **** ****** testing:

  • ***** ****: *.*.*.*.****
Comments (13)
UM
Undisclosed Manufacturer #1
•Oct 26, 2018

"Face search is included at no additional cost with Axxon Next licenses"

Are you sure about this? Information I have is that it is included in their Universe version (most expensive version), is optional in their Professional (at an extra cost making this equivalent in cost to Universe) and not available in the least expensive Start version.

It is certainly not "free" if you have to go to the most expensive version if this is the primary feature you want. I guess it would be "free" if you already needed the Universe version for some other features and got this included automatically.

See https://www.axxonsoft.com/products/axxon_next/versions.php

(1)
Avatar
Alan Ataev
•Oct 29, 2018
In reply to Undisclosed Manufacturer #1

Face Search feature introduced as part of AI pack which was released some time ago and includes features like objects tracking with TimeCompressor, forensic search by tracks, object type like human/vehicle, color and etc. Face Search as a part of our AI pack is free in Universe and optional in Professional version.

(3)
(1)
Avatar
Sergey Bystrov
•Oct 26, 2018
NetworkOptix

Axxon team, nice CPU usage results!

(4)
(3)
Avatar
Murat Altu
•Oct 29, 2018
AxxonSoft
In reply to Sergey Bystrov

3-5% per 1Mpix is not bad, actually, considering that we need also to decompress stream. 

Here is our hardware calculator to choose platform according number of channels and different VA: https://sale.axxonsoft.com/calc/calculator.jsf

But true, that for 20 cameras at just 1Mpix you need powerful server.

So, CPU consumption is currently a major issue to use this technology for say all cameras, while we make it available from license side and include to AI pack with all our analytics, like object tracker, LPR and so on.

And you know, that there are some technologies (software and hardware) coming very soon to improve CPU consumption dramatically.

(6)
(2)
Avatar
Sergey Bystrov
•Oct 29, 2018
NetworkOptix
In reply to Murat Altu

we need also to decompress stream.

Maybe just I frames, and with qsync. 

Avatar
Murat Altu
•Oct 29, 2018
AxxonSoft
In reply to Sergey Bystrov

You can do it, but quality drops significantly in this case. For best recognition we need best face captured, so minimal requirement is 6-12 FPS.

(5)
(2)
Avatar
Richie Perna
•Oct 27, 2018

As the Master Sales Representative for Axxon in the US, we have had positive feedback and response from our clients on this great technology. The Face Detection and recognition technology is second to none. We have recently deployed this technology in serval mission critical environments and are receiving excellent feedback from our partners around the country. The technology is currently being used in large K12 Environments as well as many large scale deployments.

It is critical to install the correct hardware to support this exciting technology. when setup correctly we have found incredible accuracy. We are also the Beta site for Live Face Search results including Age / Gender classification and we are having impressive results and are looking forward to partnering with Axxon to release this shortly to security community.

These videos displayed in this article are a great depiction of how the technology works and the process to set it up. Kudos to the IPVM team on a great review with detailed test results.

(8)
Avatar
George Skouras
•Oct 27, 2018

Couldn't agree more with Mr. Perna.

Being the distributors of AxxonSoft in Greece and Cyprus, we have had the chance to assist many of our integration partners/resellers in setting-up this advanced feature -among others- with great success and enthusiastic praise from the end-customers and the security managers who used it. Hotel entrances, critical checkpoints in buildings and other large-scale infrastructure like ports, shipping companies, gold mines, military facilities, to give some examples.

First of all we use the right cameras (140dB WDR IP cameras and sometimes also Super Starlight 0.0004Lux@colour) to overcome almost any extreme conditions. With the necessary study on the maximum distance (for the faces to be recognized) and the minimum resolution required to do that, we calculate the appropriate hardware for the servers (AxxonSoft provides a very handy tool to do so: https://sale.axxonsoft.com/calc/calculator.jsf). The results are amazing.

Note 1
Congratulations to the IPVM team for one more very detailed test. Fantastic work, once more.

Note 2
Regarding the bearded men misidentification, if you didn't refer that these two men are not really the same person (probably at a different point in time, with a longer beard, from a different angle and camera within the building premises etc.), I wouldn't be absolutely sure either that the faces belong to a different man. In my opinion, at this close face similarity, if I were a security manager searching for a suspect, I would surely like to have an opinion of my own and not let the algorithm decide.

@Undisclosed Manufacturer #1
This feature is indeed included in Axxon Next Professional Version as an extra cost per camera needed ("Extra Feature Package"). With this feature you can search for a face in a camera. And indeed can't be included in the Start Version.
And yes, this feature is by default included in Axxon Next Universe Version with no additional cost, but also as a "Multi-camera face search", which means that you can search for a face in many (or all) cameras simultaneously and also have the exact spots (where the faces are recognized/found) pinpointed on a 3D map. This feature is extremely helpful for the operators.

(7)
(3)
U
Undisclosed #2
•Oct 29, 2018
IPVMU Certified

Would a person who had a full beard in the reference shot, but then shaved it completely, not be recognized at all, until re-doing the ref?  

Avatar
Murat Altu
•Oct 29, 2018
AxxonSoft
In reply to Undisclosed #2

Probably, similarity will be below 90%, but definitely, even shaved, he will be found.

(4)
UM
Undisclosed Manufacturer #3
•Oct 29, 2018

I used to work for a manufacturer that placed a little white light LED on their dome cameras that flashed every second or so. It was just enough to capture people's attention and make them look up at the dome, but not so bright or fast that it was bothersome. 

(2)
Avatar
Igor Falomkin
•Oct 30, 2018
AxxonSoft

IPVM team, thank you for this thorough test.

 

The mentioned high similarity between two guys with a beard was a new fact for us. Thank you for that. I think that the reason is that we do not have enough number of men with a beard in our training set for face recognition algorithm. As for every deep learning algorithm it may be fixed easily - we are going to add more appropriate images into the training set.

 

To decrease CPU consumption we are going to use Intel OpenVINO toolkit that provides highly optimized implementation of deep learning inference engine. I hope it will let us to process more channels on the same server.

(5)
(4)
Avatar
Nabil Reigber
•Oct 30, 2018
AxxonSoft

Thanks to the IPVM team, very objective article! The latest free version of Axxon Next for four channels, which includes all the new features listed, can be downloaded from the AxxonSoft website. It must of course be taken into account that Axxon Next is a client server application and therefore we must choose its hardware platform with our appropriate calculator where we find the necessary recommendations for both servers and clients. Here is the link of our hardware calculator: 

https://sale.axxonsoft.com/calc/calculator.jsf

(1)
(2)