|Posted on December 20, 2010 at 12:50 PM|
I am afraid I have to tell you that nobody could solve the weird experiments photofit challenge. All the answers that came in (second from the right is John Lennon, the second from the left Carl Sagan...) were wrong. That was to be expected as it was exactly what Pawan Sinha had in mind when he asked an expert to produce this pictures. Here is the solution:
Bill Cosby Tom Cruise Ronald Reagan Michael Jordan.
If you know photofit pictures mainly from TV shows (as I did until I came across Sinhas paper) you will be surprised to learn how low their recognition rate is. The four faces above speak for themselves but when I tell you how they originated you will be even more flabbergasted.
The pictures were constructed with the help of IdentiKit a standard photofit software many police forces use (though I have to add that it is probably a several years old version of the program). Sinha requested an expert with several years of experience with the IdentiKit system to put together the reconstructions directly from a picture and without time constraints. That meant "the IdentiKit operator did not have to rely on verbal descriptions; he could directly consult the images we had provided him" as Sinha writes in his wonderful paper Face recognition by humans.
Let me reiterate: the operator worked directly from a picture!!! He had an image of Ronald Reagan on his desk and as much time on his hands as he wanted and the picture above was what he came up with! Reagan could have robbed banks for years if it were for photofit images he never would have been caught.
If you want to put a number on the pathetic performance of photofit images you best turn to psychologist Charlie Frowd from the University of Stirling, who did extensive research on photofit pictures. In one experiment the recognition rate was 3 percent. Again under ideal conditions this is as participants could study the pictures for one full minute before they had to describe the face on them to the person who would generate the facial composite.
And if you doubt that this results extends to reality here is a rare case where a photofit picture and the real portrait of an offender were made public. First an eye witness helped to generated the photofit picture which was printed in the papers, one day later the offenders name was identified by other means and a photo of his face was used in the search.
No wonder all the police departments I contacted do not evaluate the quality of their photofit pictures after a criminal was arrested. Most offenders were found not because but instead of photofit images!
So what's the problem? As Sinha points out in his paper the method to generate photofit images runs counter to the way humans recognize faces. Face recognition is a holistic process whereas photofit pictures are produced in a piecemeal fashion by picking out the best matching features from a large collection of images of disembodied features. But that is not how our brain processes faces. We know hundreds of people without being able to describe their eyes, nose and lips in detail.
So what to do? Charlie Froud, Peter Hancock and Vicki Bruce developed a way of generating suspect likenesses called Evofit that fits our brain much better than picking out isolated eyes, noses and mouths. Working with Evofit the witness chooses from 72 random faces those 6 that are most similar to the offender. The computer then mixes the facial features of those faces to produce 72 new ones and the witness chooses the 6 most similar again. After three such rounds Froud achieved a recognition rate of around 25 percent.
Considering the shortcomings of conventional photofit programs I thought it would be fun to hold a photofit not look alike contest. If you find published photofit pictures and the real mug shot send them to firstname.lastname@example.org. I will publish them here.