Slashdot Mirror
Virtual Robots Fooled By Visual Illusions
Roland Piquepaille alerts us to research out of University College London in which virtual robots, trained to "see" as we do, were duped by optical illusions the same way humans are. Here's one of the illusions the software system fell for.
That's the idea. The robot sees how we see, and thus fell for the illusion of green and orange, when both squares are green.
no, it's not a difference in lighting.
the central squares are in fact the same color on your monitor, (pretty close to hex: 647316).
this is very similar to this famous color constancy illusion.
I think that's not what the parent meant.
Lightning change through the day, so the actual color of reflecting objects also is changing. But the object didn't physically change and your brain "corrects" color, that is abstracts them (you wouldn't say your blue car to be blue the day and dark gray the night, it's simply blue).
In the illusion at hand, left sphere is interpreted as being lit by a red light, while the red sphere is interpreted as being lit by a blue light.
Of course, "Ceci ne sont pas des sphères", only pixels, so the comparison on interpreted colors fails.
But given two real spheres, lit one by a red light and one by a blue light, and reflecting like in the picture, the colors would be different with these two spheres lit by a same white light.
I have discovered a truly marvelous proof of killer sig, which this margin is too narrow to contain.
What's your point?
If you put a white tile in a room with red light, do you want the robot to see a white tile? or a red tile?
If you change the light, from red to blue, do you want the robot to still be able to identify it as the same tile?
If yes, then this should be considered "a difference in lighting" and not an error on the part of the computer vision algorithm.
...that's a difference in surrounding lightning.
...WHEN those colors are ISOLATED. (i.e: putting the name "red" "orange" "purple" on a color you see alone).
Human visual system (as most other senses) work not by absolute values (i.e.: it doesn't see that the color '#c0ff20' or whatever), but mainly by comparing the signal with signals from the surrounds.
Thus what we technically see is that on the left object the central case looks much more "greener" than its surrounding, in the right object, the central case is "much more orange" than the surrounding. In fact, when the mask is enable, the colours do change from the point of view of the visual system : we were seeing contrast with two different surrounding, now we see a contrast with a third surround (mostly black). We see three different contrasts, even if from the computer's point of view the color is them same (the same RGB triplet / same intensity on your CRT/LCD)
If the scientist are trying to build efficient visual systems, they are probably mimicking this "works-by-comparing" method that the nature is using.
That's why we can recognise the same object, during day, during night, with weird lights, displayed on the screen (worse colour gamut) or on a print out (even worse color range). Because the relative difference stay the same, even if the colour as-seen-by-a-computer change.
The same is valid for any other sens, or in fact, any other information that is processed by neurons. Everything works by comparing (across several signals, across time, etc.). There's no such thing as "an absolute value" in the information carried by neurons.
That's also why all those "but the human eye can only x thousands of colors" (usually mocking the latest 32bit, 48bit, floating point or whatever color depth), are fundamentally wrong.
Yes, the human visual system can only distinguish a hundred or so colors....
When two colors are put next to each other, the human brain can suddenly distinguish much more subtle variations (each color would be considered as "brown" when seen alone, but next to each other, you can use thousands of different shade of brown and the eye will still see the difference).
That's also why radiologist are fond of high contrast / big depth screens : because all those difference in shades of grey *can* be distinguished and *are* revellent for the diagnosis when displaying X-Ray pictures.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
Of course, "Ceci ne sont pas des sphères", only pixels, so the comparison on interpreted colors fails.
That's what makes it an optical illusion. Your brain is interpreting visual information based on a context which causes a failed interpretation. That could be a definition for "optical illusion".
These aren't colored spheres, and no one said they were colored spheres. It's just an arrangement of colored patches, arranged in such a way as to give your mind a bunch of visual cues that there are different colored lights shining on those patches, causing your brain to misjudge the actual color of those colored patches. Hence, it is an illusion.
The general explanation, as I have seen it given many many times previously, is that, rather than write a /. story which links to some science/tech article, roland will paraphrase the article in his blog, and link the /. story to his paraphrase. This is a means of gaining ad revenue for himself and his employers (ZDnet, I think?), but it doesn't give any ad revenue to those who actually did and wrote up the research.
Is this true? I don't know. I never RTFA.
SIGSEGV caught, terminating
wait... not that kind of sig.
It's not about lighting at all. It's about lateral inhibition. The neurons in your brain, when connected in a nural network, send inhibition signals to their neighboring neurons. Areas close on you retinas have neurons on the same part of the brain, so the squares surrounding the center square are going to activate neurons close to the ones you use to see the center square. Now the lighter squares send a stronger signal to the neurons surrounding the center square, thus sending MORE inhibition, thus making the square appear darker. w0rd.
I wonder if you're getting a kind of Persistence Of Color effect.
Here's one my favs:
http://www.johnsadowski.com/big_spanish_castle.php
Cube On! (http://stores.ebay.com/PuzzleProz)