Jonathan McDowell wrote:

However, I think the statement that I was making is nothing to do with
your very valid concerns. My point is that we have integer pixel numbers
(P1,P2,...) like (241,32,81) and fractional (real-valued) pixels (X1,X2,...)
like (241.3, 32.1, 81.9). Different software systems use different
schemes to map between these; I have seen both
[A] Xi = Pi Pi = (int)Xi
and
[b] Xi = Pi + 0.5 Pi = (int)(Xi-0.5)
and of course this is an independent choice from the schemes
[F] i = 1,....N
[C] i = 0,....N-1


People may prefer the latter although I cannot see why, but the latter
is WRONG. If people use it then they will LIE to the recipient about
the WCS of their data. The fact that the center of the VOXEL (we are
2 dimensions a lot) is the reference point of the VOXEL is part of the
WCS papers and has been approved at all levels by the FITS committees.
It was certainly understood from the very beginning of FITS and
explicitly stated at least some places. The argument about this that
appeals to me most is to think if the VOXEL in real space as a cube.
The only point in that cube that does not change under rotations is its
center. In all other cases, if you for example transpose an image, you
will change the "natural" (think left side) reference point of every
pixel on that axis. Since one cannot change the reference point of
projective geometries without regridding the image, this is an
unpleasant concept.

Guys - the correct way to clarify some misunderstandings in the written
standard needs to be determined. BUT, the issue of reference point was
resolved long ago. Let us not reawaken it and destroy the good work we
have done to convey coordinates with data.

Eric Greisen