[gdal-dev] Overview scales

[Even Rouault]

> Yes, with interpolation methods that take into account more than a
> single pixel of the source image, this effect will not be that obvious.
> But that doesn't resolve the problem if one is interested in more than
> just visualization, but e.g. in deriving positions from the image
> content, as in least-squares image matching at decreasing image pyramid
> levels and/or between images with largely different scales.

Imagine that you have 101 source pixels and create an overview of 50 pixels.
If you use linear resampling, to compute the first ovr value we will take into 
account the corresponding first 2 source pixels plus a tiny bit of the 3rd 
source pixel, since the spatial extent of the target pixel overlaps a bit that 
pixel. So the contribution of source pixels is evenly distributed to target 
pixels, which seems to me to be the adequate solution, rather than deciding 
that to use 2 source pixels for each target pixel, except the last one.

> > So yes there's a bunch of code change, and also test suite "golden"
> > results to be updated. That'd need a champion.
> 
> Unfortunately, I'm surely no GDAL champion. So I hope that one of those
> picks up this problem!

I meant someone to champion the effort, ie do the work ;-)

> 
> Cheers,
> Willi.
> 
> > Even

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com