🚀 go-pugleaf

Bryan Morse writes
> As I understand it, splatting is a technique for resampling images.
> Imagine each original pixel as a snowball and throw that snowball
> at the new image ("splat!").  Each snowball as it hits distributes
> its snow (intensity) over a region.  Each new pixel is then incremented
> by the contribution of that snowball to that pixel.  Usually, the
> distribution is Gaussian, but others have been used.  Once you've
> thrown all of the snowballs and added their contributions, you have
> a new resampled image.
>  
> Basically, the idea is to reverse the usual interpolation.  Rather than
> computing each new pixel as a function of the old pixels, compute the
> relative contribution of each of the old pixels (their splat) to the
> new ones and add these together for the new pixels.

This can work provided that the "size of the snowball" is correct for
the output grid. If the output grid is sparse relative to the input grid
then it should work fine with lots of small snowballs - small intensities
and small area (suggest adding based on original intensities and then
normalizing/scaling values). If the output grid is dense relative to the
input grid then the snowballs should have large area (to ensure every
output grid point is covered) and large intensities (one snowball
"splats" many grid points).

This snowball scaling aspect seems critical. The projection of the input
grid points onto the nearest output grid points is known to leave holes
and create weird collision effects. Fixing it by making the input grid
points distributions (snowballs) and projecting the distributions onto
the output grid in an additive manner may or may not be faster than a
classic warp with interpolation. Has anyone done a study? A timing
comparison with real code?

I would be most interested in seeing an algorithm statement, either in
a posting or by reference to a paper (or book?).

Mark R. Thomsen