Fix bounding box "rounding"

[larrybradley]

This PR is a followup to #81, which used the round function. We don't want to use round to compute the bounding-box edges because of its behavior at half-pixel values. For such cases round (in Python 3) rounds to the nearest even number, e.g. round(0.5) = 0.0 but round(1.5) = 2.0. That behavior may make sense in a statistical sense, but it's incorrect here for computing bounding box edges.

In this PR, I use what I implemented in my refactor of photutils.aperture:

imin = np.floor(value + 0.5).astype(int)
imax = np.floor(value + 1.5).astype(int)

which "rounds" consistently for the bounding box edges.

Also note that the above behavior changed for Python 3. In Python 2, round would always round away from zero, e.g. round(0.5) = 1.0 and round(1.5) = 2.0. This is yet another inconsistency that we want to avoid (and another reason why round should not be used).

[cdeil]

@larrybradley - Thanks!

You put this:

ixmin = np.floor(xmin + 0.5).astype(int)

Is this the same?

ixmin = int(xmin + 0.5)

If yes, I think I would slightly prefer it.
But in any case, just a thought ... we'll leave what you think best.

Another question: should we really have a test for 0.5?
It's a fractional float. Is it really guaranteed to round the way you have in the test on all machines?

[larrybradley]

You put this:

ixmin = np.floor(xmin + 0.5).astype(int)
Is this the same?

ixmin = int(xmin + 0.5)
If yes, I think I would slightly prefer it.
But in any case, just a thought ... we'll leave what you think best.

For non-negative xmin they are the same, but they can differ for negative xmin, e.g.

>>> xmin = -1
>>> np.floor(xmin + 0.5).astype(int)
-1

>>> int(xmin + 0.5)
0

Negative values in the bounding boxes can occur if the region is only partially on the image. The Mask methods handle those cases when applied to data.

[larrybradley]

Another question: should we really have a test for 0.5?
It's a fractional float. Is it really guaranteed to round the way you have in the test on all machines?

I'm not actually sure.

[cdeil]

I'm not actually sure.

+1 to remove that test and then merge.

[larrybradley]

@cdeil Done!

[cdeil]

I'm merging this now.

@larrybradley - Should we leave the 0.5 rounding example in the docstring or also remove that?
We can always take it out in master

[larrybradley]

@cdeil I removed the 0.5 rounding example in master.