How Much Has Facial Recognition Improved In 2018?

Rafael, thanks for the question. A few thoughts:

Many companies have been doing / trying to do facial recognition for many years. Even back in 2005, there were a number (note: I was at 3VR where we got $50 million investment to do this - poorly).

That noted, I do agree with you that there has been a spike in the past two years of companies offering facial recognition solutions.

In terms of what is being offered for facial video surveillance (i.e., use surveillance cameras to match / alert on people 'in the wild'), right now, it seems to be mostly a handful of niche companies and a number of companies inside of China (3rd grouping is home cameras which I am excluding since they are not designed for any professional scale of surveillance or monitoring).

We'd like to test something that might work. I am not sure right now which though considering most are small players and the rare positive feedback I hear is generally not specific or detailed enough to build confidence.

Surely, facial recognition will continue to improve but doing it in adverse lighting conditions (bad inputs) and against large groups of people (many people look a like) make it a very challenging application in production.

It's important to understand that facial recognition ain't facial recognition.

What suffices as a second factor of authentication does not suffice as a primary authentication factor.  A 99% "hit rate" here could mean that 99% of the time it will accept the correct person bearing the qualification, but it may accept the incorrect person if they look a bit like the correct person 10% of the time (see some of the mother/son father/son/daughter examples on the iPhone on YouTube).

When using FR to blacklist customers etc., a 99% hit rate may mean that they alarm on the correct person 99% of the time, but they may alarm on the incorrect person who looks a bit like the correct person a few percent of the time (or maybe one in 1000 innocent customers).  Not too bad if you have only one or two people on the blacklist, not so good if you have 100 or 1000 people on the blacklist.  Mind you, that's still better than my own human FR algorithm which has real trouble matching face to name in my grey matter database.

At present FR (despite the claims of the current vendors) is still in its infancy, and is best used as an adjunct to other methods (like 2 factor authentication, guards tracking suspicious customers on CCTV, etc.)  By all means, experiment, but I wouldn't encourage clients to rely on FR as their primary means of identification, an oracle for blacklisting, replacing a PIN or signature to approve payments, etc.

Hi Alf,

Long time no see. I completely agree with you: one have to separate FaceReco for Access control and FaceReco for Blacklisting users (Stadium, Retail, Airport, ..)

John,

Regarding Face recognition technology, its' true that its Performance and Consistency and its Price has not yet reached the ones of the best leader in Biometrics for access control => 2D Fingerprint. It is widely known by specialists that ~95% of the Biometric readers that were shipped in 2017 are ... Fingerprint readers. Then Face then Iris then the other HandShape/Fingerpvein/HandVein and other methods fill in the remaining 5%. Maybe that number will be 94% or 92% in 2018. But that does not change the big picture.

If you want to confirm that ... I advice that you purchase IHS Markit report (access control) at least one time. I believe it gives a reliable trend based on major manufacturers globally.

https://technology.ihs.com/Categories/450443/access-control-fire 

I believe that FaceReco for Blacklising users is the field of CCTV HW/SW manufacturers (Axis, Genetec, Milestone, Hik, ..), and I have no expertise to share on that field. 

FaceReco for Access Control is very different. Here, you have to consider two cases:

A- Very high secure areas (Data-Center, Gov, Airport) => in that case most user hold a credential (DesFire EV1 or HID iClass or HID Seos or Passport, Nat_ID_Card). So it is quite easy to process a "card then face" 1:1 (verification matching) with very high hit rate for that use case. Note that at Suprema Inc, we use FaceStation2 (Infrared sensing technology). That is a 2D technology based on ~30x Infrared templates (including evolution templates). The only limitation so far is that Face templates cannot be stored on Smart Card due to size + need for evolution (because your face changes .. look at your pics from 5 years ago ^^). Note that Suprema FaceStation2 has this face change machine learning + supports most Cards (EM, HID Prox, Mifare/DesFire EV1, iClass SE, Seos, NFC, BLE) + all formats (I mean the wiegands 25~64 bits) + HID/AES/DES Encryption (for the Secure ID).

B- Bio-Secure Areas that need FaceOnly recognition. Here the end-user request is "I want security AND convenience". That means users wish not to use cards at all and no contact skin contact at all (that's a fair request in some cases). So what is requested here is a 1:N (one to many) face matching. For that use case Suprema limits the faces to 3K users. Not because of a problem of device memory size, but because we have found out that real FAR/FRR (false acceptance / false rejection rates) and matching time are only acceptable up to that point. I insist on real FAR/FRR because you sometimes can find FAR/FRR in bio-devices specs and that is a biggest "BS" ever made by Bio-manufacturers: FAR/FRR given are theoretical (not based on real faces but just mathematics values based on algorithm). True thing is that Real FAR/FRR will depend on face chances (makeup, hairs, beard, environment, tiredness, gained/lost weight, face attitude/smile_or_not), Face Subject Angle (and here for Disablities Act), DB size (100 users or 2000 users?) and some other factors that shall remain confidential.

As a sum up:

Based on 4 years of installation of FaceStation and FaceStation2, I am very confident with A- case above with unlimited users (=1 million) and B-Case up to 3,000 Faces (one user can have two/multiple faces = with/without glasses).

One point in John discussion above: CPU processing power plays for a lot in performance because it allows to make use of heavier Algorithms/tricks. Here in FaceStation2, Suprema uses a 1.4Ghz Quad-Core.