Terminology for Automatic Image Alignment
aruzinsky
via simultaneous adjustment of rotation, size, and location? I know this
is a type of two point registration thing, but I am unsure of the exact
terminology for googling references to automatic algorithms.
2. What would be considered fast for rotation between +-45 degrees, resize
between 0.75 to 1.5X, and location between + - 25% of the image dimensions,
to within single pixel accuracy, for a 3.3 mp image?
pixel.to.life
1. The most common term, I guess, would be 'image registration'. It is
usually considered a multi-variable optimization problem. Depending on
what optimizer you choose, all or some of the variables (in this case
3 rotations and three translations for a rigid transform only) might
be perturbed at a time.
2. Do you know some correspondence between the images a-priori? E..g.
do you know the location of at least 3 landmarks in both the images?
If yes, you can try aligning pointset to pointset (assuming the bodies
they are on stay rigid). Search for a paper by Besl and McKay I
believe on google...on shape registration. If you dont know the exact
correspondence, you can still pick a finite number of landmarks
automatically from both images (but on the same object) and employ an
ICP like method to align two pointsets. I doubt if you will get sub-
pixel accuracy this way though.
If however you dont have this info, you can still try registering the
images using statistics based methods. Look for 'mutual information
based image registration'. You can choose from among a number of
metrics depending upon the nature of images etc. For a good
combination of metric and optimizer, you might be able to hit sub-
pixel registration accuracy.
Hope that helps....
On Mar 25, 12:34 pm, "aruzinsky" <aruzin...@nospam.general-
aruzinsky
Thank you.
"Do you know some correspondence between the images a-priori?"
Only that the images are of the same scene taken with a hand held camera.
The purpose of the digital alignment is to make stacking practical for
photographers without a tripod. You can't expect typical photographers to
do manual landmark registration.
serg271
Other term to google is "Markerless tracking"
You can use cross-power specter - it was discussed in this group
already.
One of the links
www.ist-matris.org/publications/WIAMIS-AR.pdf
To my shame I had forgot how it's working during last discussion and
made some nonsense remarks.
Don't pay attention to them if you looking through group archives.
I'd realized the errors of my way by now and successfully implemented
method in my code too.
The method is taking Fourier transform, multiply at the conjugate of
Fourier transform of the second image,
normalize and make inverse transform. The spike is the shift of the
images.
After that align images, convert to polar coordinate and repeat
process to get polar offset, which is the angle of rotation and the
scale.(Fourier-Mellin transform)
The problem is, in my expirience 25% translational shift is borderline
for simple single-spike form of that method. With shift around 30-40%
spike is blurred/noised and some more sophisticated Fourier analysis
is needed, what exactly - I have no idea. Of cause you can make
shortcut - make rough approximation (possibly with the same method)
and apply method on already roughly aligned cut-out images. There is a
modification of this method with sub-pixel correction, but it seems
don't help much.
On the positive side the method is very fast for featureless method -
only three FFT.
I also didn't test rotational form, but it should have less problem
than translational because there is no windowing problem.
Please tell us how will it go if you will use that (or other method) -
I'm doing related thing (Augmented reality)
serg271
>
> On the positive side the method is very fast for featureless method -
> only three FFT.
altogether + some lesser ops for interpolation to polar coordinates
and multiplications/normalizations of first stages.