Quividi’s Digital Sex Change Feature
Manolo Almagro, Q Division Managing Partner
Last week’s DSE 2009 show gave me a chance to get a quick look at 3 companies that offer real-time, automated audience measurement solutions for the digital-out-of-home industry, Quividi, Cognovision and Tru-Media. Apparently though I was more of a man at the Tru-Media and Cognovision booths than via Quividi, who told us that “they don’t do Asians” after successfully demonstrating to us some (rather rude) gender re-assignment.
Manolo (male), Adrian (male), and Quividi staff (female)
Each of these software solutions does the same thing (just, in slightly different ways).
Each of them use on-site cameras to process images, then analyze data (in real-time) and record the information in the form of; number of viewers, how much time each viewer was facing the screen and viewer demographics (sex/age). But before you start waving the privacy issue flag – rest easy, none of these systems store images – just the numerical data.
That explanation makes all this sound rudimentary – but facial recognition analysis isn’t easy, in fact it’s takes a whole heck of a lot of effort to get it to work accurately. To better appreciate the sophistication of real-time image processing, here’s a 5 second primer on how some facial biometrics software works.
For each face, the software analyzes at least 80 facial “landmarks” of a viewer; to name a few – the distance between the eyes, width of the nose, depth of the eye sockets, shape of the cheekbones, length of the jaw line… blah, blah, blah – you get the picture.
In any case, the analysis happens every millisecond and concurrently for every person that registers in the camera’s field of view. The software then executes a comparison of data and makes an approximation of the viewers sex and age. – in most cases, the typical accuracy rate of these systems is around 90% – especially when it comes to discerning the sex of the viewer.
The biggest challenge that all these vendors face is the issue of ‘race’ – the fact that each nationality has unique facial characteristics.
What may be considered a ‘female’ characteristic for a Caucasian does not bode well for, err… let’s say, a middle aged, Filipino male (that would be me, see here)
So when faced with a multiple ethnic crowd of viewers, how can they discern sex?
Case in point, take a look at the photo at the top of this post. Adrian and I visited the Dynascan booth which happened to be demonstrating the Quividi solution.
Looking from left to right – you’ll see me, then Adrian (pointing and laughing at me!) and an unidentified female, this beautiful French lady was the wife of Quividi’s Chief Scientific Officer, Paolo Prandoni, Ed who was working the booth.
Note the colored indicators (circles) around our faces. Red / Pink = female, Blue = male.
Needless to say we tried a few times to recalibrate the accuracy of the Quividi software BUT alas my gender reassignment was permanent!!
The software was convinced that I was a female (thus our sex change comments). When we asked the female booth attendant why this was happening, she told us “we don’t do Asians” – perhaps not a good idea to employ your wife on the stand then?, Ed
Last time I checked (and I do it often) I am a male!
March 2nd, 2009 at 12:09 @547
¿All systems detect viewers at 125º angle vision?
TruMedia do it, it´s important to know the real audience.
Roi Iglesias
March 2nd, 2009 at 13:14 @593
Dear Roi,
I think it’s good to be honest and explain that you have a partner with Tru-Media. So, you don’t try to give a comment as an objective opinion, if before you don’t explain the reality.
Dennis
March 2nd, 2009 at 15:43 @697
Dear Dennis
Many thanks for your reply. If you have more information about other audience measurement systems I´ll apreciate to comment it with you.
I comment about TruMedia´s solutions because I test some systems and from my point of view it´s important know all the audience (as sure as possible).
If you have a comparison of systems I´ll love to meet you. My cuestion on this post is because I´m curious to know the benefits of other audience measurement systems.
In Spain we have very few cases of audience measurement in place.
Best,
Roi Iglesias
March 10th, 2009 at 09:00 @417
Quividi would like to address the technical aspect of this post.
Quividi’s gender recognition algorithm has a success rate in excess of 86%, as openly stated in our website and in our documentation. What this means is that, on average, one person out of ten will be either misclassified or not classified at all. We absolutely want to stress that any classification errors do NOT depend on ethnicity. As a matter of fact, these errors are not easy to trace back to simple facial features since the algorithm does not work the same way a human brain would. (In passing, note that humans score “only” 95% on gender classification tests!)
The 86% success rate is achieved via a training procedure on a database of over ten thousand manually labeled faces. The face database has been built as to include the largest possible distribution of age brackets, ethnic types and hairstyles and has no “racial” bias. We sell our product all over the world and we need consistent success rates. At the same time, we have never developed nor will ever develop any ethnicity classification mechanism in our software — that would simply be out of line with Quividi’s core values.
You should remark that, as opposed to our competitors, we are the only audience measurement company to show gender results in real time in our demo. We can afford to do this because we are very confident about our software and because we have always been completely transparent about the capabilities of our product. Normally the occasional and unavoidable mistake is more of a source of chuckles than of a serious critique and we are sorry that this was not the case. Don’t take it personally and do not feel too special: you are just one in ten.
March 10th, 2009 at 12:20 @555
Oliver – I appreciate the fact you took time to comments. However, Its important to reiterate that the basis of the posting was derived from a specific comment from someone representing your product – when I asked why I, a male, filipino was being recognized as a female, I was told that the system was not calibrated for “asian faces.”
September 28th, 2015 at 20:59 @916
[…] being brought to market to provide some proxy for measurement (of traffic, eyeballs, and with some occasional gaffes, gender) in a digital signage environment, it seems to me that the angst of the agencies is NOT […]