Can you recognize the faces below? (Answers at bottom of post). The faces are compressed so the width is only 25% of the original photo.
 |
| Source: Pawan Sinha, MIT |
 |
| From Wired |
But apparently the human visual system does not depend critically on exact measurements. Scientists have discovered recognition doesn't suffer much with proportionally distorted faces. As long as the measurements are proportional within a region or across a single dimension—that is, as long as the ratios are preserved—recognition performance isn't greatly affected.
So what structural aspects of the face are the most important for recognizing faces? The researchers conclude: "It is possible then that human encoding of faces utilizes such ratios (we refer to them as iso-dimension ratios), and this might constitute a useful strategy for computer vision systems as well. "
Credits:
G. J. Hole, P. A. George, K. Eaves, and A. Razek, B. Effects of geometric distortions
on face recognition performance, [Perception, vol. 31, no. 10, pp. 1221–1240, 2002.
Pawan Sinha, MIT
Answers: Ronald Reagan, Jason Alexander, Prince Charles, George Bush, Robin Williams, Woody Allen
Credits:
G. J. Hole, P. A. George, K. Eaves, and A. Razek, B. Effects of geometric distortions
on face recognition performance, [Perception, vol. 31, no. 10, pp. 1221–1240, 2002.
Pawan Sinha, MIT
8 comments:
Orientation is a critical variable in recognition. Turn the faces upside down.
Interesting. I wonder if our brains have a harder time recognizing faces that are compressed vertically instead of horizontally?
I'm trying it myself in photoshop right now, but I can't tell, since I already know what all the faces are. It would be interesting to take another set of photos of other celebrities, compress them vertically, and see if people can recognize them.
With the same degree of compression, could we distinguish between an older and younger Jason Alexander? Could we spot Woody Allen without his glasses? If not, it may have less to do with proportion and more to do with signature characteristics that survive extreme manipulations such as in caricature drawing.
This is really interesting. I have been watching videos on Youtube about how to draw using the Bargue technique, and it starts off with finding the ratio of what they call the "Notional Space Box" which is the box that would fit the outermost points of a figure. They make a big deal about getting the ratio of the box exactly right before the rest of the drawing is done. But you can take a photo and radically change the proportion of that "space box" and it's still completely recognizable. So now I'm wondering if that's really the best way possible to try and get a likeness. I wonder if there is way to incorporate this fact to make a detailed and sequential technique similar to the Bargue technique, but where the ratio of the "notional space box" is not important.
Robert, the Barge method, as it is often taught, focuses on accurate observation, accurate measurement, and accurate transcription to the drawing. This can result in a superficially precise drawing that often lacks character. In the French academies, "character" was a goal highly prized by the best masters, and it was a quality distinct from measurement or precision.
From what we know now from the modern science of face recognition, the question of recognizing faces hinges on identifying how a given face varies from the norm. The artist's job is to accentuate those variations, and that may be the path to achieving character. This is at the heart of what Walter rightly describes as "signature characteristics" which stand out even if certain proportions are distorted.
Keith, I don't know if those signature characteristics would survive a stretch in the opposite dimension, but I suspect they would. They also would survive a substantial degradation of the image through blurring or pixellation, and I'll get into that in a separate post.
Colonel, you're quite right that turning the image upside down makes a face much less recognizable, and so does inverting the values, and I'll cover that one in the future as well.
Thanks James for sharing your thoughts, and for this really interesting post. That's a great point about "character". This fact shows how right they were, that exact and mathematical correctness in shapes is not the end all, be all of portraiture.
Keith, it seems that portraits that are stretched horizontally are a little bit harder to recognize than portraits stretched vertically, for the simple fact that the portrait is usually already taller than it is wider, so the likeness seems to survive longer being stretched vertically than it does being stretched horizontally.
I'm going to have to look more into this, I think this is a really important factor in portraiture. Thanks for this post James.
Facial recognition has always fascinated me. I'm particularly amazed by an aspect of it that was not addressed in this post; namely, our incredible ability to recognize a person from childhood that is now middle aged or older.
Imagine instantly recognizing an old classmate at a 50th high school reunion that you haven't seen since you were 18-years old. Think about how extraordinary that is!
How does our brain not only utilize iso-dimension ratios to recognize familiar facial proportions, but also understand how those proportions may change over time through aging?? Black hair now white, bushy eyebrows, numerous new wrinkle lines, softened features, new glasses, receding hairline, and yet recognition happens in an instant! Astonishing!
Jim, your comment reminds me of the sequence in Woody Allen's Annie Hall where the kindergartners talk as if they were grownups. I also think of the famous documentary film "Seven Up" which was updated every seven years, allowing us to see faces change and personalities develop. I've had the opposite effect happen, too, where I see one of my sons' school age classmates years later as a teenager or an adult, and I have no idea who they are—especially if some years have gone by in between. Growing up can be the ultimate disguise.
Post a Comment