Lease this WebApp and get rid of the ads.
Daniel Buck
John Cleese, Amelia Earhart & Doc Holliday
Fri Jan 11, 2019 04:38

John Cleese, not only a brilliant comedian but a sharp thinker, did a four-part series for BBC, "The Human Face," that touches on human and machine facial recognition, and he revisits the subject in his new book, Professor at Large: The Cornell Years (2018). Both the series and the book might be worth checking out for those interested in the facial-recognition topic.

A point he makes, salient to our discussion, is that humans are better at recognizing faces than machines, that is, the disguises and other distractions that fool FRT don't as easily fool humans. That's why FRT is always best combined with human analysis. Machines, obviously, are faster at viewing massive image files. No doubt about that. Artificial intelligence software can swiftly trawl a 250,000 image file, and come up with ten possible matches (though we never know how many it missed, e.g., the Boston Marathon Bomber case), but then an actual human being -- you've heard of the species? -- sits down to look over the images (and check their background records) to see if among them is the person they are looking for.

In the typical contested Old West photo case, however, we are only looking at a couple of photos to begin with, usually just the photo of the actual person and the photo of the possible match. Humans are much better at this kind of two-person line-up, the one vs. one analysis, however you want to describe it. Humans with some knowledge of the Old West celebrity in question are even better, in part because they have in their head years of knowledge about the facial characteristics of the celebrity, and in part because they can get into non-facial factors, drawing on their own independent knowledge, which are as important if not more important, for example, history, provenance, photographic processes, and chronology. The simple example is that Alec Baldwin is not Millard Fillmore in spite of their resemblance because they lived in different centuries (to the best of my knowledge). A photo taken by a Shanghai photographer in the late 19th century who looks like Will Rogers is not Will Rogers because he was never in Shanghai (that I am aware of). Or an 1872 CDV of a man who looks very much like Butch Cassidy is not the outlaw because he was only 5 that year (last I checked).

A classic real-life example is that it took a Japanese researcher half an hour on a library computer to determine that a supposed 1937 photograph of a captured alive Amelia Earhart had been published two years prior to her flight in the region. And so on. In short, humans might be failures at many things, but facial recognition and related context is not one of them, or are not two of them, depending on how you parse that sentence.

Another commonplace example is that you can recognize your neighbor, whom you've seen many times, a block away by contour and gait even though you can't make out the face. Squirreled away in your brain are hundreds if not thousands of bits and pieces of visual memory of that person that come flashing together in a nanosecond. The "blink" of Malcolm Gladwell fame. It annoys hopeful flea-market photo finders no end when a Doc Holliday student takes a brief look at their possible treasure and says, "nope, no way." That blink is a brief look, plus 20 years of research.

Cleese -- I think it was Cleese -- said that when we went upright from scrambling around on all fours we moved from dependence on our nose to relying more on our eyes. What he really means is the eyes plus the brain. Dan

Click here to receive daily updates