Slashdot Mirror
Facebook Thinks Occlusion Is the Next Great Frontier For Image Recognition
An anonymous reader writes: Researchers at Facebook AI Research (FAIR) have published a paper contending that image recognition research is now advanced enough to consider the problem of occlusion, wherein the objects AI must identify are either partially cropped or partially hidden. Their solution is the predictably labor-expensive route of human annotation of existing image-set databases, in this case 'finishing off' occluded objects with vector outlines and assigning them a z-order. This article looks at the practical and even philosophical problems of getting IR algorithms to 'guess' objects usefully, and asks whether practical IR research might not be currently limited both by the use of over-specific image datasets and — in the field of neural networks — by problems of theory and limited 'local' processing power in critical real-time situations.
Facebook does conduct research into AI. They need such technology to more effectively mine their vast database for advertising information.
Occlusion handling is the difference between 'subject identified as Joe Bloggs' and 'Subject identified as Joe Blogs wearing Adidas trainers and posing in front of a Skoda. Increase targeting of well-known fashion brands, decrease targeting for automotive products.'
But...
Will it be susceptible to optical illusions?
Everything I write is lies, read between the lines.
you
It will have to be. The ability to figure out what you're looking at with incomplete information is exactly what leads to optical illusions, you can't really have one without the other.
A bullet may have your name on it but splash damage is addressed "To whom it may concern."
If a series of images is available and observer or target or intermediate objects are moving, occlusion will vary image to image and the nature of the delta portions should be highly informative for recognition. This requires an object/region re-identification subsystem.
Also, scene context statistics should be used, much as preceding utterances are used in speech recognition. Given that we've already recognized a situation type with this that and the other object-type in it in this (possibly dynamic) relation, what are the a priori probabilities for these other types of objects to occur in scene, and assess occluded objects against highest probability objects in situation type. Much more constrained/determined recognition problem in which pieces of objects might suffice to identify them.
Where are we going and why are we in a handbasket?
Will it be susceptible to optical illusions?
Vision systems based on artificial neural nets are susceptible to many of the same optical illusions as people, and for mostly the same reasons. The basic vertebrate eye has been around for 530 million years. If optical illusions were easy to avoid, nature would have figured out a way to do it by now.
So if regular object recognition is such a solved problem, why to they need people to manually prepare the images? I'd just take a normal image, recognize the objects, and then partially cover some of them to train their algorithm.
> Especially black faces
This! I cook a lot and post pictures to Facebook. It can never find my face, but it thinks my stovetop is a face.
The use of vector completion and all is a good idea, but it seems systems like that would work better in conjunction with other techniques, like trying to consider context of the area where you are in. What is behind a tall narrow object varies a lot depending on if you are in a jungle vs. a parking garage...
"There is more worth loving than we have strength to love." - Brian Jay Stanley
That sure makes sense. Don't tell them, though. The inability of image recognition software to handle cropped pictures is one thing which my better replacement for CAPTHCA depends on. CAPTHCA sucks because humans aren't much better at computers at recognizing squiggly letters. We are, however, MUCH better at recognizing certain specifc types of images when they are cropped and rotated.
Pull back. Wait a minute. Go right. Stop.
Enhance 57 to 19. Track 45 left. Stop.
Enhance 15 to 23.
Gimme a hard copy right there.
Thank you, I am starting to work on countermeasure right away so I can keep my private life. I'll arrange so it thinks I am some politician, a giraffe, an SUV or something else. There is all kinds of illusionist shows on TV so it shouldn't be that hard.
Everything I write is lies, read between the lines.
Thanks for the idea. I'm going to go right now and arrange two fried eggs and a strip of bacon in a smiley face and post it as my Facebook profile picture.
You are welcome on my lawn.
Example:
https://i.ytimg.com/vi/7I95IFw...
You are welcome on my lawn.
What about a kind of genetic algorithm to evolve candidate 3D models, and the model that best matches observations and context "wins". However, that is computationally intensive. But, it is highly parallelizable.
Table-ized A.I.
In fact, this won't stop at merely recognising faces that are partially obscured - in the not so distant future, they will be able to recognise faces that are completely absent!
Funny thing is that virtually all AI vision systems have problems with black faces. It isn't human racism that is the cause or 'machine' racism, its the physics of cameras and optics and light itself. At least with modern HDR cameras it is a problem we have some hope of beating..
Below the speed of light Special Relativity is one of the most accurate theories in physics - above the speed of light..