[ptx] PTmender update

[Daniel M. German]

Hi everybody,

Just a quick update. I am "almost" done with the colour-brightness
adjustment. It is "almost" because it is still untested. Unfortunately
I don't have reliable Internet access, but I'll try to do a CVS commit
as soon as possible (I fixed a couple of bugs this week). So far
PTmender can stitch TIFF_m, but it can only generate TIFF.

The colour-brightness code is almost complete (I should it end today)
and I'll start testing. 

Please, one thing I really need is for somebody to run the example
from Helmut's web site
(http://www.path.unimelb.edu.au/~dersch/cbcorrect/cb.html) and send me
the debug.txt (I don't have a machine that can run PTStitcher). That
way I can compare my output with his output.

With respect to the algorithm used I can roughly describe it as
follows (of course this is my interpretation):

Compute all the histograms of all the intersections of photos, n *
(n-1)

Start with reference photo as corrected.

While there are photos to correct
  Find new photo with the largest overlap with corrected photos
    Compute Histogram of overlap of new photo with corrected photos
    Compute a curve that will match the photo to correct histogram to
     the correctedPhotosHistogram 


This will result in an array of correction curves (actually, each
intersection has 6 curves, see below) that is applied to each
photo. These curves are discrete (arrays of 256 values).

PTStitcher computes the histogram of each channel, their average,
red-blue and red-green.

BTW, two photos are not corrected if the overlap is less than 1000
pixels.

Currently PTStitcher corrects only 8bit images.

In my opinion theses algorithms should be implemented standalone
(separate from the stitching algorithms).


 Pablo> Dear Philippe,

 Pablo> Thanks a lot for the program. I haven't had time to implement this approach
 Pablo> myself yet. Actually, some people are recreating PTStitcher from scratch, and
 Pablo> as far as I know this algorithm is the main missing ingredient (its not inside
 Pablo> libpano12 as far as I know). Do you mind if I forward your mail to them?

 Pablo> Actually, I have planned to add a similar feature to hugin/nona. My experience
 Pablo> with the correction from PTStitcher was that for panoramas with many images the
 Pablo> correction did introduce color casts. Probably due to error propagation. I also
 Pablo> think that vignetting might influence the correction result. Hopefully the new,
 Pablo> soon to be released vignetting correction in hugin/nona will also improve this.

 Pablo> Ideally I'd like to use an algorithm that estimates all adjustments at the same
 Pablo> time. For this a regularized estimation might be useful, which punishes high
 Pablo> deviations from the original images (maybe only in userselectable images).


--
Daniel M. German                  "I come out of traditional mathematics
                                   where we prove things,
                                   but we don't charge people
   Donald Knuth  ->                for using what we prove. "
http://turingmachine.org/
http://silvernegative.com/
dmg (at) uvic (dot) ca
replace (at) with @ and (dot) with .