[ExI] ai application, was: RE: suv hipsters

[spike]

Artificial intelligence software can be stupid as all hell and still qualify
as AI.  Bumper sticker version:

 

AI need not be intelligent, only artificial.

 

I have been having a lot of fun trying to figure out a car-recognition
algorithm and I think I have one.  It is a memory hog, it is stupid and
slow, but at least it's artificial.

 

I have thought of an algorithm for car recognition for which I invite wise
counsel from code hipsters and Stanford AI classers.  We take 12 pictures of
a chariot, every 30 degrees, with front view being defined as the 12 o'clock
and the right side view being the 3 o'clock.  Then we get a bunch of
volunteers to take 12 photos of their ride and collect all.  Then we
separate them into a collection of 12s, 1s, 2s, etc, so we have 12 files,
each with photos of various Detroits taken from a similar angle.

 

That's step 1.  Then we take a photo of an arbitrarily chosen auto at an
arbitrarily chosen angle.  Then, I write an algorithm to recognize the
wheels, more specifically the rim elipse.  We take the ratio of the vertical
apparent diameter to the horizontal (minor diameter) to estimate the angle
at which the picture was taken.  For instance, if the minor diameter is half
the major diameter, then we know it is either a 1, 5, 7 or 11 o'clock photo.


 

Then we look to see if it has either two red or two white features above and
to one side, indicating either headlights or tail lights.  The white
features to the right for instance indicates a 1 o'clock view, two red
features to the left of the rear tire indicates it is a 5 o'clock view.

 

Then once we get an angle, we start comparing the test chariot with the
collection of photos donated by volunteers.  We JPG-ize both, then take
horizontal rasters, looking for high contrast patterns in common.  The high
contrast areas would be the roof columns perhaps.  The software would be
simple enough, even I might be able to write it.  It *might* be able to pick
the one from the collection that is the best match.

 

Think that would work?  

 

Have we any JPG gurus here?

 

spike

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20111023/0ea98b66/attachment.html>