[ptx] Enblend Multiresolution Spline Blender

[Andrew C Mihal]

On Sun, 7 Mar 2004, Pablo d'Angelo wrote:

> Hmm, seems to take quite a while on these big images.. ;)
> Actually I'd like to incoporate the blending stuff into nona. The seaming
> lines could also be calculated from the info created during the remapping,
> like the angular distance, instead of thinning the masks.

Yeah, the thinning is really slow. But it is general enough to work on 
overlap regions with very complicated shapes. If you have a mathematical 
model of the outlines of the images, there are faster ways to get the 
transition line. There are lots of Medial Axis Transform papers that talk 
about this. I haven't read these.

> It might also be worthwile to calculate the seaming lines with the image in
> mind. I thought about using a search for a seamline with minimum cost on a
> cost image formed by the overlaps. (it could be the a simple difference
> between the overlaping parts (of the greylevel, maybe the hue channel)
> , the correlation of small patches).

The cost metric should take into account both stitching irregularities and 
distance to the edge of the overlap region. Some experimentation will be 
required to find out how to weight these factors.

I had an idea to modify the thinning algorithm to also thin the mask away
from areas where there are stitching irregularities. Then the transition
line would be as far away as possible from both the edge of the overlap
region and stitching errors. I tried it but it did not come out the way I
expected. I was able to identify regions in the overlap zone that are
closer to stitching errors than anything else. But these regions
completely overwhelmed the information about distance to the edge of the
overlap region. So I could not decide whether these regions should be
filled in from the left image or the right image. You really need to have 
both distance-to-stitching-artifact and distance-to-overlap-edge 
information.

Andrew

---------------------------------
Andrew Mihal
www-cad.eecs.berkeley.edu/~mihal
mihal at eecs.berkeley.edu