Prev: How to insert a blob data that is image into sql table in matlab file
Next: bar3 BaseValue
From: bhavin on 1 Jul 2010 07:25
hi

I have determined the least squares similarity transformtion (includes scale rotation and translation) between two color images using some technique. Now i am faced with the task of finding how accurate the transfomation parameters are. Right now I am comparing my transfomation matrix to that calculated in MATLAB using the cpselect tool in which i manually select a pair of correponding points between the two color images and then compute the similarity transfomation. I need to do so for over a 1000 pairs of images and since the cpselect tool is manually performed, the comparison task is so time comsuming. So i would like to find out if there is an automated script in matlab or any free ware that is known (ie paper published on it or is proven to be the best or at least one of the best techniques) to accurately find the similarity transfomation between two color images.

Will really appreciate your help
From: ImageAnalyst on 1 Jul 2010 09:46
There is no way to have some automated way to tell cpselect what
points to pick. You have your automated way (which doesn't require
cpselect) and the manual way (which is cpselect) and you want to
compare how well your automated way matches various images to their
target "perfect" orientation with the manual method using cpselect.
I'm guessing you're hoping your automated way will be closer and more
accurate than the manual/human-chosen way. Why is there no way to
automatically tell cpselect which points to pick? Well let's say
there were. How would you choose those points? Well you'd have to
have some kind of automated image analysis to try to pick reasonable
points. Well that's what your automated method already does. And
once you have the set of points, you can just call maketform and
tformfwd and there's no need for cpselect then. Yes manually calling
cpselect a thousand times is tedious and time consuming but after a
day or two you'll be done.

Why not just use the PSNR to compare the images? You'd have to crop
out the non-common regions of the image near the borders, but then for
the parts of the image that are supposed to overlap, the PSNR should
give you a good metric to see how well different transformations
worked.
 | 
Pages: 1
Prev: How to insert a blob data that is image into sql table in matlab file
Next: bar3 BaseValue