Biometry application

CGAL/cgal

Biometry application

April 1st, 2012

Dear CGAL developers, Dear CGAL users,

In order to put an end to rumors about GeometryFactory selling CGAL to Facebook for a biometry application, let me give you the background, and let me explain why we think that what we do is not evil.

Just as nuclear energy can be good or bad, biometry is neutral as such. The particular application we contribute to is about matching lip prints on cigarette butts with Facebook users who threw the butts away. This allows people to act, who are fed up with cigarette butts lying around virtually everywhere: in front of the Statue of Liberty, as well as along trails in Yellowstone National Park.

One of the goals of social networks is social engineering, that is making mankind better. The workflow of the biometry application is
1 that you take a picture of a cigarette butt lying at a place where it should not,
2 that you upload this picture to FB together with the GPS coordinate of where you shot the picture, so that FB can identify the culprit by matching the lip print on the cigarette butt with the extensive database of faces.
3 that FB tells him or her that this is not a well educated behavior (smokers in fact know that it is not nice), and puts repeat offenders on a wall of shame.

As it is technically rather fascinating, let me just tell you why it is not a trivial problem. We can't reveal all the details as we filed a patent. Lips leave prints, just like fingers do, but fingers tips are flat compared to lips, cigarettes are cylindrical, the max curvatures of lips and cigarettes are mostly orthogonal where they touch. Just as maps are distorted towards the poles, so are the photos of lips and cigarette butts. Additionally, only a small part of the lips touches the cigarette, only one side of a cigarette butt is visible, and it needs several uploaded photos of FB users to get a high fidelity lip model. To get a grip on the orthogonal curvatures and distortion we throw in the anisotropic surface mesh generator Jane Tournois works on, and we implemented the Lipschitz distance estimator to get the confidence value for the lip-cigarette-butt match.

FB opted for CGAL, as the exact computing paradigm helps to avoid false negatives (FB has to avoid accusing people who not even smoke). The floating point filters are key to not only being exact but also fast, as there are millions of smokers who have to be identified. Obviously, FB combines that with yellow-finger detection, the timeline information they have ("Hey, I am in Venice"), and whatever other data available.

As I wrote in the beginning, biometry is not evil as such, and improving society seems a more moral goal to us than target advertising. Similar applications we have in mind, are the detection of counterfeit Dolce and Gabana glasses, Lacoste T-shirts, and Louis Vuitton handbags. FB Singapore pointed out that identifying chewing gums on the sidewalk would interest them, but it is not clear yet how to adapt the Lipschitz distance estimator.

It was a mistake not to have discussed this at the last CGAL developer meeting or openly in the forums, but I had not expected such strong reactions when some of you heard from Jane that we work on "some biometry stuff".

best regards,

andreas
--
Andreas Fabri, PhD
Chief Officer, GeometryFactory
Editor, The CGAL Project