XPost: sci.image.processing, comp.graphics.algorithms, comp.ai.neural-nets   
   From: broeker@physik.rwth-aachen.de   
      
   [F'up2 cut down to single group --- should have been done by OP!]   
      
   In comp.graphics.algorithms Jack Zoken  wrote:   
      
   > My company's business has a need to recognize automobiles (make and   
   > model) from an orthorectified aerial image. The imagery is 6"   
   > pixels, 24 bit color.   
      
   Let's see.  6 inches per pixel gives you about 26 x 16 pixels for a   
   car of 4x2.5 meters in size.  Less if you're not looking down at it   
   vertically.   
      
   I *strongly* suspect there's strictly no way at all you can pull that   
   off.  Measuring distances to 6 inches simply isn't precise enough to   
   distinguish shapes of objects that are more-or-less all of the same   
   general size.  And that's before you even begin to worry about things   
   like how the lighting (glare spots --> over-exposed image, edges lying   
   in the shadow --> no contrast) affects your ability to even tell where   
   the exact edge of that car is.   
      
   You'll probably be able to tell an old VW beetle from a Lexus or Jeep,   
   but you'ld have a *very* hard time telling, say, a New Beetle from the   
   classic one, or a Lexus from a Mercedes.  You can almost certainly   
   forget about being able to tell different models of the Mercedes "E   
   class" or "S class" from each other.   
      
   In other words: you will need better input or it's an impossible   
   mission.   
      
   --   
   Hans-Bernhard Broeker (broeker@physik.rwth-aachen.de)   
   Even if all the snow were burnt, ashes would remain.   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)