BBC Featured Deep Learning Face Recognition Analyzed (Digital Barriers)

By: Brian Karas, Published on Aug 30, 2017

The UK's largest broadcaster, BBC, recently featured Digital Barriers (also based in the UK) on a segment highlighting face recognition technology. The TV segment showed Digital Barriers' SmartVis Face [link no longer available] being used to monitor public spaces for appearances of "terrorist watchlist" suspects.

Digital Barriers claims the use of deep neural networks is extending the capabilities of their SmartVis Face product, allowing them to achieve new levels of performance and accuracy.

This report analyzes the SmartVis product, based on conversations with Digital Barriers, and outlines where the company says the product has made advancements, and where limitations still exist in facial recognition.

*** **'* ******* ***********, BBC, ******** ******** ******* Barriers (**** ***** ** the **) ** * segment ************ **** *********** technology. *** ** ******* showed ******* ********' ******** Face [**** ** ****** available] ***** **** ** monitor ****** ****** *** appearances ** "********* *********" suspects.

******* ******** ****** *** use ** **** ****** networks ** ********* *** capabilities ** ***** ******** Face *******, ******** **** to ******* *** ****** of *********** *** ********.

**** ****** ******** *** SmartVis *******, ***** ** conversations **** ******* ********, and ******** ***** *** company **** *** ******* has **** ************, *** where *********** ***** ***** in ****** ***********.

[***************]

SmartVis ********

******** ** ******* ********' ***** analytics ******* ****. ******** Face [**** ** ****** available] ** *** ******-*********** component, ***** ** ***** on ********** ******* **************** **** ************** ** ****. ***** **** acquisition, *** ******* **** **** have **-********** *** ******** of *** ********, ********* introducing * **** ******** basis ** *** ****** to ****** ****** ***** in ***-******* **********.

******** **** ** * server-based ******, ********* ******* from ********* ******* ** real ****, *** ********** alerts ** **** ***** on ************* ******** *** matching (** *** ********) a **** ******* * database ** ******** *****.

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

******* ******** ****** **** have **** * **** learning ******** ** ***** a ****** ******* ** recognize ***** ** ******** distinctive ******** **** ************* various *****. **** ** in ******** ** *********** systems ***** ********* **** to ******* ********** *** distances ******* ******* ********** of * **** (*.*.: spacing ******* ****, ******** between **** *** ****, etc.).

*** **** ******** ******** allows **** ** **** a *****, ** * lack ** * ***** faster ** ******** ** previous ********. **** **** enables ****** ******** ***** when *** **** ** the **** ** ************* different **** *** *****(*) used ** ****** ** in *** ********.

* ***** **** ** sampled ****** ******* ****** of *****, ******* ** from **** * ****** "best ****", ***** ******* Barriers ****** **** *********** to ********* ******** ** the ******, *** *** speed ** ********, ** the ****** ***** **** multiple ***** ** *** face, ********** ********* **** data *** *** ******** algorithm ** *******.

******* ******** ****** ***** capabilities ***** *** ******* well ****** *** "***-*********" checking, ***** * ****** may *** ** ***** a ****** *********** ****** is ** ***, *** they *** *** ***** to "******" **** *** system ** ******* ******** at * ****** ** passing ******* * **** obvious ***** *****.

Field ** **** **************

*** ********* ***** * minimum ** ** ****** between *** **** (~*** PPF / ~*** ***) for ******** ***********, *** **** see ******** *********** ************ up ** ** ****** ******* the ****, **** ****** numbers (*.*.: ** ******) not ******** *** ********** benefit ** ********.

** ** * ***** in *** ***** ** view *** ** ************** analyzed.

***** ***** ** ** theoretical ******* ********** *********, the * **** ***** and ***** ************ ********* lead ** ********* ***** 1080p cameras.

Enrolling *****

***** *** ** ******** in *** ******** **** static ****** ******** ("********"), from * ****** ******** to * **, ** from ** **** ***** segments. *******, ******** *****/****** of *** **** *** captured *** ******* ******** in ******** *** ******/**** later.

Server ************

******** **** **** ** an *****/******* ***** ******** (i5 ** ** *********), and **** **** ** accelerate *** **** ****** network *********. ******* **** are****** *** **** ******** cards.

***** ****** ************ **** vary ***** ** *** number ** ********, **** of *** ******** ***** database, *** *****/****** ***** desired ** *** ********. An ******* ****** **** a ****** *** **** card is ********* *** * streams, **** * ******** of ~*,*** ***** *** return ***** ******* ** real-time (GPUs *** ** ***** to ****** **** ******* simultaneously, *** *** *** CPU ************ *** ******** for ***** ********* ** applications **** **** **** throughput ** ******.

*******

***** ****** **** ~$*,***-$*,*** per *******/******, ********* ********* on *** **** ** the ****** ******** *** customer ***** ** ********. The $*,*** ***** ** most ******, ********* ** Digital ********, *** ********* running ********** ** *** range ** ******* ******* to ~*,*** *****.

Sales *******

**** **** ***** ******* *******, ******* ******** ***** the ******** ********* ******* standard ******** ********** ******** globally.

Software - **** ******** *** ** ********

***** ******** **** **** include * "**********" **** interface, ********* ** ******* Barriers, **** ********* *** the ******* **** ** an ****** **** ****** a *** ***** **** interface, ***** ***** ** a ***, ** **** other ***********-******** ********. ******* Barriers **** **** ************ with ***** **** ********* and *******, ** **** as **** **** *********, and *** **** ****** customers **** ****** ***********, if ******, *** ********** feeds. *** ******* ****** that *** ** *** variety ** ***** ******** base ***** ****** ** no "********" *******, *** that ***** ****** ****** some ****** ** ****** integration ** **** ***********.

*** ******* **** **** are ******* ** * new ******** ******* **** will **** **** ************ in *** *****-***** ******* to **** ** ********* to * ******* ******** base.

Upcoming ********

******* ******** **** **** the ******* ** ********* non-face ******* ** ** development *** ****** **** logo-tracking. **** ** **** intended ** ***** ** an ********** **** * face, *** **** ** to ***** ****** ** track **** *********** ** things **** *********** ** TV *****/******** ******, ** clothing **** ** ******, etc.

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

**** *********** ******* **** faced **** ********** **** the *****. **** *** technical: ************ *** *********** much ****** **** ******* surveillance ****** ******* *****, requirements *** ******* ********* of *******, *********** ** deal **** **** ***** at ****. ***** ****** are ******* ******** **** citizens, ** ****** ****** like **** ** ******* obscuring ***** *** ********** the ******* **** ******* a ***** **** ** analyze. *******, **** *** made *** ******* ****** to ******, *** ********** for ******** **** * person ** * ********* will ******** ** ******** if **** ** ****** in *****.

**** *** *** ****** any **** *********** *******, primarily *** *** ******* outlined *****, *** ****** comment ** *** ******** Face *** ******* ** real-world *********** ** ***** systems, ** ** *** own ************ *****.

Other **** *********** *********

******* ******** ** *** alone ** *** **** recognition ******** *****. ********* such **********,*****,******* **** ********* ******** face *********** ************ ** core ** ***** *********. Additionally, ***** [**** ** longer *********] *** ********* also **** **** ********* face *********** ************ ********.

** *** **** *** particular ******* **** * given **** *********** *******, please ***** * ******* detailing **** **********.

Comments (15)

Christian Laforte

08/30/17 08:13pm

I tested four or five solutions the past 10 years. They were disappointing. At best I got some of them to work reliably when doing a 1-to-1 matching, i.e. between an access card and the photo previously assigned to it. They never worked "in the wild". Mismatches were common.

This past month we have installed NEC Neoface in our office. I'm really impressed. It took a few weeks (part-time) to learn the system and tweak the many settings, mainly to accommodate some of our older, lower-resolution cameras, and running 10 cameras on an old spare server. It's now working reliably as long as subjects generally look toward a camera. NEC seems to err on the side of caution, i.e. it would rather not generate a match unless it's absolutely sure this is the right person. They have an impressive track record - some customers have databases with millions of people.

BTW organizations seriously considering face recognition should read the US NIST face recognition vendor test.

https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt

John Honovich

IPVM | In reply to Christian Laforte | 08/30/17 08:16pm

Christian, good feedback!

It's now working reliably as long as subjects generally look toward a camera.

Well, what do they recommend there? It's hard to make people look towards the camera in a real world, uncontrolled surveillance application?

Also, any sense of how well or poorly it scales up? Btw that I mean, it's easier to get good results with a small watchlist and a small number of people being surveilled but put it in a large scale application with tens of thousands of different people passing daily and hundreds of people on a watchlist and it's a lot harder.

Undisclosed Manufacturer #2

In reply to Christian Laforte | 09/05/17 04:03pm

Facial recognition accuracy has improved greatly over the past ten years, so it's no surprise that you're more pleased with a more recent offering.

While FRVT is important, a more appropriate test is the recent "Faces in Video Evaluation" test. https://www.nist.gov/programs-projects/face-video-evaluation-five

Christian Laforte

08/30/17 08:35pm

> Well, what do they recommend there? It's hard to make people look towards the camera in a real world, uncontrolled surveillance application?

Two approaches that would be cheap and effective:

- Put a monitor with animated content right next to the camera, that the person is likely to look at. For instance, an animated menu (for a restaurant), a welcome message (for corporate reception), a live playback of the surveillance camera (in a retail environment to deter theft), a quote-of-the-day for a church, live sports (in a bar) etc.

- Put the camera next to an intercom, a door bell, or other device that a person will naturally face when interacting with it.

> Also, any sense of how well or poorly it scales up?

I haven't tested it yet but the system appears well designed to scale to millions of subjects. Their technical documentation is extensive and gives lots of details on how to achieve it. From a performance standpoint their software services architecture is quite similar to ours (that scaled to 100 servers for one big project). As for the accuracy, the NIST report is best suited to answer your question.

> Btw that I mean, it's easier to get good results with a small watchlist and a small number of people being surveilled but put it in a large scale application with tens of thousands of different people passing daily and hundreds of people on a watchlist and it's a lot harder.

Yes, that's what the US NIST was tasked with testing, e.g. for airports, border protection, etc. false alarm rates could easily get out of control and make a multi-million dollars systems useless. I've seen it in person 4 years ago at Dubai airport where a rack full of face recognition servers was permanently turned off. (I won't mention that vendor.)

Personally I'm only aware of NEC and Safran/Morpho having a solid track record in these kinds of large government contracts. Of course this is a somewhat controlled environment. A few months ago there was news that a police in UK was able to arrest a criminal through street surveillance. That was NEC, and that's why I became interested in trying it out myself.

John Honovich

IPVM | In reply to Christian Laforte | 08/30/17 08:40pm

Thanks for the feedback. Interesting to hear about the scaling.

- Put a monitor with animated content right next to the camera, that the person is likely to look at. For instance, an animated menu (for a restaurant), a welcome message (for corporate reception), a live playback of the surveillance camera (in a retail environment to deter theft), a quote-of-the-day for a church, live sports (in a bar) etc.

I laugh at this a little bit because I have heard this advice many times before but it's generally hard to accomplish that type of 'look at the birdie' technique. It, of course, depends who the subject / watchlist / adversary is but the type of professional criminal that are often on such lists tend to be savvy about not staring up at potential cameras, welcome messages, etc.

Phil Shanahan

In reply to John Honovich | 09/12/17 04:57pm

I would imagine that mounting cameras much lower than normal would help with this. I'm thinking a corridor or entrance door where the camera is mounted to a wall at about 1.2 meters high looking directly at the target. As people look in front of them the camera should easily see their face. Typically cameras are mounted on, or close to the ceiling which would be more like 2.4 meters. not sating mounting cameras at 1.2 meters is practical but I don't see any other way to get people to 'look' directly at the camera

John Honovich

IPVM | In reply to Phil Shanahan | 09/12/17 09:09pm

Phil, agreed, mounting cameras lower will help. The #1 objection to that is vandalism (see Camera Vandalism Statistics).

The secondary issue is that when the camera is mounted lower people block others more easily. A higher mounted camera has the benefit to literally look 'over' people and that's a practical issue as well for face recognition if you have a scenario where multiple people will regularly be walking by simultaneously.

Christian Laforte

08/30/17 08:57pm

> the type of professional criminal that are often on such lists tend to be savvy about not staring up at potential cameras, welcome messages, etc.

You're right, detecting hardened criminals in the wild isn't full-proof yet and won't be for many more years. But I think there are other business opportunities that can live with less-than-perfect accuracy, e.g. retail.

Undisclosed Manufacturer #3

In reply to Christian Laforte | 09/06/17 06:42am

If you go through some uk airports these days, there are cameras installed in the passport control areas with an animated ring of leds around the lens. These are obviously designed to make you glance up and look directly at the camera, its basic human nature.. Anyone purposely evading could be easily pulled aside for 'further questioning'.

If only I could find a link to them..

John Honovich

IPVM | In reply to Undisclosed Manufacturer #3 | 09/06/17 12:02pm

uk airports these days, there are cameras installed in the passport control areas

I believe that and, in those environments, you have the law and authorities to enforce that immediately. Where it is harder is in less controlled environments like a mall or a store or a city street, etc.

Undisclosed Manufacturer #3

In reply to John Honovich | 09/06/17 01:00pm

Agree 100% with you John. But that's the difference between a soft and hard target.

Undisclosed End User #1

08/30/17 11:46pm

Disclaimer, I worked for manufacturer for almost 10years that did Face Analytics and have cut my teeth in the school of Hard Knox, I once drank the Kool-Aide from the firehouse as well.

Welcome the world of Behavior Analytics in uncontrolled environments with uncooperative subjects. I suggest you not use "real-time alerting" as a knock-out requirement it will never work as planned w/o specially placed sensors and subjects who are in lane of analysis while looking in the direction of the sensor for a minimum of 2-3 seconds. Even with all of that there are so many variables.

Best value for FACIAL SURVEILLANCE is for forensic searching, you don't have real-time matching issues. You accept all of the false matches is search and weed them out but cannot do the same in real-time. Just like "G" Search for Best BrewPub, you take all the crap it gives you and filter out what you don't want to find the best cold beer, rank indexing. Take it for what it is, a SEARCH TOOL!

Only place where it works well enough is where there are controlled environments and subjects. Why does it work well in mantrap, well Johnny knows he must stare at the specially placed Face Camera until it matches him, he is a static subject and its easy for the algorithm to do its thing with less pixels and computing resources.

Why does other Face Recognition systems like those at DMV or pokey work well matching? Simple answer when you limit all of the variables then the analysis is much easier

Static Subjects told to stand still until image is captured
Camera adjusted to same eye height for all subjects input
FoV is the same every time
No hats/glasses/smiling
Environmental - there are lights and a backdrop
Pixel Density is the same for every image, often very high resolution

Want a tip to beat Face Detection, look down when you walk in the area of analysis.

John Honovich

IPVM | In reply to Undisclosed End User #1 | 08/30/17 11:50pm

Good points.

Want a tip to beat Face Detection, look down when you walk in the area of analysis.

And criminals generally know this, even without face recognition or deep learning, they know it's better to look down so cameras don't get a clear shot. And / or to wear a hat or sunglasses, etc. It does not take a criminal mastermind to figure that out.

Undisclosed Manufacturer #2

In reply to John Honovich | 09/05/17 04:06pm

Of course, if all criminals were masterminds, fingerprint identification systems would never have gained traction. Good point, though.

Brandon Knutson

IPVMU Certified | 10/17/17 09:03pm

Fascinating concept! "Dubai Airport is replacing security checks with face-scanning fish"