In a new History Channel documentary, image recognition experts say that a recently discovered photograph suggests that Earhart and her navigator, Fred Noonan, survived a 1937 crash in the Pacific and were later photographed on the Marshall Islands.
A blurry photo reportedly taken around the time of Earhart and Noonan's disappearance shows someone seated on an oceanside dock and another person standing on the dock, to the left of the seated person. Numerous other people also appear in the photo, some standing, others in sailboats moored at the dock.
An image recognition expert who studied the photograph said the torso measurements of the seated person match those of Earhart, and the hairline of the individual standing to the left matches that of Noonan.
CNN, which aired a July 5 story about the documentary, spoke to experts who challenge that claim, notably Dorothy Cochrane, curator for the Aeronautics Department at the Smithsonian National Air and Space Museum. She told CNN that she has seen the photo featured in the History Channel documentary and called it "interesting" but "not definitive."
"People take photos and interpret them, and they're free to do that," she told CNN. "It has not persuaded me."
Would Cochrane and other skeptics be more easily persuaded if the analysis had been done by a machine?
We hear considerable hype these days about the power of machine learning and, specifically, image recognition. Are the image recognition services available today capable of delivering on their promise?
To cut through the hype and gauge the value these services offer, CapTech recently tested the six leading image recognition services: Amazon AWS, Microsoft Azure Computer Vision, Clarifai, CloudSight, Google Cloud Vision, and IBM Watson. We recently completed a white paper that summarizes the results.
For the tests, we presented each service with the same set of approximately 4,800 images, distorting many by blurring, overexposing or underexposing, positioning the images at odd angles, and otherwise recreating real-world conditions. We then evaluated nine distinct functions (including facial recognition) and the services' accuracy and confidence in their answers.
The results of the facial recognition tests were particularly interesting. The services we evaluated were often stymied by distortion and low resolution of the images. An example of this involved simple inversion of the image. The services recognized actor Brad Pitt in an upright image, but did not recognize him when the image was upside down. The image purported to contain Earhart and Noonan are extremely distorted and grainy.
The exception to this was CloudSight, which could identify celebrities despite heavy distortions and filters. The probable reason for CloudSight's performance isn't better algorithms. Rather, CloudSight appears to rely on human labor to recognize images.
As for Amelia Earhart: we already know that human experts disagree on the question of who appears in the History Channel photo, so CloudSight probably couldn't resolve the dispute. And given the inability of other image recognition services to correctly identify an upside down Brad Pitt image, I doubt that these services would fare any better in recognizing Amelia Earhart or Fred Noonan in a blurry photo taken on a dock 80 years ago.
Maybe technology will one day give us a clear and undisputable picture of Earhart's fate, but we haven't reached that destination yet.