Pittsburgh Post Gazette, PA
03-25-07
Why police composites don't always hit mark
New system for police sketches takes into consideration that witnesses typically recall the entire face, not individual features
Sunday, March 25, 2007
By Mark Roth, Pittsburgh Post-Gazette
If the FBI sketch of the man who abducted and raped a 12-year-old Wheeling girl last month is a good one, it may lead to an arrest and conviction.
But the odds are that it won't be, say scientists who do research on the accuracy of composite drawings.
Laboratory tests around the world have shown that whether composites are created by a professional police sketch artist, as in the Wheeling case, or by using one of today's sophisticated computerized systems, they have an accuracy rate of only about 20 percent -- and that's when the witness describes a face shortly after seeing it, says Charlie Frowd, a psychologist at the University of Stirling in Scotland.
Under more realistic conditions -- the witness waits anywhere from a few days to a week before helping to prepare a composite -- the accuracy rate plunges into the single digits, Dr. Frowd said.
The low matching rates have very little to do with the skill of the artists, the power of the computerized facial systems or how hard the witnesses are trying, the researchers say.
Instead, the problem is that most composites are put together by asking a witness to describe the parts of a face -- the eyes, nose, mouth or chin -- and then assembling those pieces to create a likeness.
The popular FACES computer composite system, for instance, offers witnesses 63 head shapes, 361 types of hair, 514 eyes, 593 noses and 561 lips to choose from.
But most of us don't recognize or recall faces based on their individual features, said Gary Wells, a psychology researcher at Iowa State University who has been active in trying to reform eyewitness identification techniques.
Several brain studies have shown that we tend to see a face as a whole, and we pay more attention to the relationship among the parts of a face than we do to the parts themselves.
"Every cognitive scientist who has studied faces has concluded that faces are processed holistically. In fact, we now know that at least as early as six months of age, babies are engaged in the holistic processing of faces, not individual features," Dr. Wells said.
"You can take people who've been married 15 or 20 years and the husband or wife can be quite incapable of describing a single feature of the spouse's face accurately," added Christopher Solomon, technical director for a British composite company called VisionMetric Ltd.
For that reason, Drs. Wells and Solomon believe most of the computerized composite systems used today have serious flaws.
Many police officers who prepare composites would disagree.
"All I can go with is the results I get, and they are very good," said Blase Kraeer, a Pittsburgh police detective who has worked with the EFIT computer composite system for six years. "It's very rare, maybe one out of 10, that the composite hasn't been good enough to make the identification of the person."
Actual examples of composites and photos of those later arrested can show a mixed result, though.
Detective Kraeer is justifiably proud of a composite of a suspect in a triple homicide at a Homewood diner in January 2002, in which an 8-year-old girl was one of the victims. The penetrating eyes on the masked composite are a close match for one of the men convicted in the killing, William Thompson.
But another example shows how problematic composites can be.
After a series of brazen sexual assaults on women in the East End in late 2002, Pittsburgh police issued an EFIT composite of the suspect, based on several witnesses' descriptions.
Yet the photo of Daniel Wesley, the man who was later arrested and convicted of the crimes, based largely on DNA evidence, is markedly different from the composite. His face is much broader, his eyebrows thinner, his ears smaller and the spacing of his features much closer than in the composite.
Detective Kraeer said that he generally won't release a composite if the witness description isn't strong enough, but much of the time, it is a person's ability to remember key features of a face that makes a composite good.
"There are certain features they absolutely remember," he said, "and we can get those almost perfect."
Change coming slowly
But researchers say the very act of describing features can interfere with a witness' recollection.
Dr. Wells said there is evidence that when a witness constructs a composite, that image can start to replace the memory of the actual face the person saw.
Jonathan Schooler, a psychology professor at the University of British Columbia in Canada, calls the phenomenon "verbal overshadowing." In experiments he conducted, witnesses who described a face they had seen had more trouble picking it out of a lineup of similar faces than witnesses who didn't provide a description.
In the real world of police investigations, though, the scientists' skepticism has not made many inroads.
The most widely used computerized composite systems today are still based on the "parts to whole" approach, and most of the tinkering done on them in recent years has been aimed at making the drawings more lifelike and offering an ever-increasing number of features.
William Scigliano, the executive who oversees FACES, said he has no statistics on the accuracy of his firm's composites.
"But I can tell you we saved a few lives and I'm really proud of that -- and how do you quantify that?" said Mr. Scigliano. "My comment would be we caught a serial rapist in Florida, and a serial rapist in Fresno; we have some great success stories.
"In the end," he said, "this is just a tool; it's not a panacea."
But a tool needs to be based on the right concepts, said Dr. Frowd, the Scottish researcher, and the systems that depend on the internal features of a face are emphasizing the wrong things.
The whole-face approach
That's why Dr. Frowd and Dr. Solomon at VisionMetric are both developing new composite systems that use a more holistic approach.
Dr. Frowd's system is called EvoFIT, and is about to be tested in several British police departments.
EvoFIT starts by generating 72 male or female faces -- 18 at a time on a computer screen -- and asking the witness to pick the six images that are closest to the face in his memory. EvoFIT then "breeds together" those six faces to create 72 more faces, and the witness is asked to go through the process again.
By the third round, the witness is asked to pick the one face that is closest to the one he remembers, and that is the one used as the final composite.
The two most important parameters in EvoFIT, Dr. Frowd said, are the overall shape of the face and its texture, not any individual features.
The system has been tested in one criminal investigation, a serious sexual assault case in southern England. The EvoFIT composite was a "powerful image that caused the victim distress," he said.
In lab tests, the holistic approach has performed about three times better than any of the "parts" systems, he added.
VisionMetric's EigenFIT system uses a similar but slightly different approach.
The witness looks at nine faces to start with, picks the best one, and the system produces nine more. But in EigenFIT, the witness can refer to a particular feature that isn't quite right, such as the nose or the shape of the face, and the system will incorporate that information into the next set of faces, Dr. Solomon said.
By tweaking the system in that way, the witness can arrive at a good composite after about 12 sets of faces, he said.
"It's very Darwinian really," he said. "The idea is, you're kind of favoring these combinations of facial components that you implicitly recognize are more suitable. And the great thing is you don't have to verbalize it."
Keys to a good composite
Suni Chapman, who sells EFIT in North America and does much of the training of police officers who use it, is skeptical about whether the holistic systems will be that much better.
"How can they claim their systems have such a higher percentage when they haven't caught a bad guy yet?" she said.
Ms. Chapman, who has done many composites herself, believes the skill of the person interviewing a witness makes a major difference in the quality of a composite.
For one thing, the interviewer cannot suggest any of the features by saying things like "Did he have a long nose?"
The interviewer also has to let the person recount the crime at his own pace. "The person's mind goes from A to N to L to Z. You note down things but you don't interrupt. You just come back to them later."
These open-ended methods help a witness re-create the image of the suspect's face, she believes. "It's like you're trying to shake the cobwebs off a dream."
Ms. Chapman and other composite artists note that their drawings are only one part of the investigative process, and police and prosecutors then have to develop other evidence against a suspect to convict him.
In the Wheeling case, the FBI sketch artist feels confident that his drawing of the rape suspect is a strong likeness, and it already has generated more than 600 tips, said Wheeling Police Chief Kevin Gessler.
But Chief Gessler also noted that composites are just "a tool to get you started. It's one step along the investigative process."
Iowa State's Dr. Wells suggested, however, that in some cases, composites could lead to the arrest of the wrong person.
"To the extent that the composite is used to find a suspect, it can in certain cases beg the question of whether the person looks like the suspect or like the composite."
He noted that in 75 percent of the cases in which convicts have been exonerated through DNA evidence, the original convictions were based on false eyewitness identifications.
"I'm not trying to say you should never use composites," Dr. Wells said, "but to note that there are real potential negative consequences of using a composite and getting a negative result."
He believes many police investigators understand that risk.
"The greatest nightmare a cop has is to be going after the wrong guy. Any time he's going after the wrong guy, it means the bad guy is getting a free ride."