[ImageJ-devel] ROI -> Overlay bug

[Stephan Saalfeld]

Hi Michael,

you are of course perfectly right.  So the PSF is elongated along the
scanning direction.  However, that PSF would still be symmetric, except
if the sensor has a tendency to collect photons only later or earlier
during exposure.  So to make my statement clearer, a point sample taken
from the physical world is almost never a point-sample, but it's
representative coordinate is most likely not at the top left corner of
the area between samples.  If the PSF is symmetric, that coordinate is
at the center of the sample.

Take a pixel sensor of a single channel CCD camera.  That sensor has an
area to collect a number of photons.  Be it square.  Then it looks as

x-------+
|       |
|   o   |
|       |
+-------+

In that picture, it is more than obvious that the intensity collected by
that sensor has it's 'location' not where the 'x' is but where the 'o'
is.  So, from that perspective, it is much easier to think about
coordinates centered in the pixel.

Alternative example:  If you rotate an image of 3x3 pixels, coordinates
are

(0,0), (0,1), (0,2)
(1,0), (1,1), (1,2)
(2,0), (2,1), (2,2)

by say 90 deg clockwise around it's middle pixel, whose coordinate is
(1,1).  Then doing so with the top,left-model moves that pixel to (1,0).
Not very convenient.  The center-model would keep it in place as is the
natural expectation.  You may now argue that the center of that image is
at 3/2=1.5, but that requires that you know that the 'size' of the last
sample is 1x1.

The situation is different when it comes to display, where all AWT comes
from.  On a display, pixels are fixed size rectangles, and in this case
the top-left model has some benefits.  E.g. when it comes to scaling:
you can scale a pixel image by 1/2 by combining 2x2 pixels into a single
one.  With the top,left-model, that would correctly scale the coordinate
of a pixel by 1/2, with the center-model it would shift the pixel by
0.5px down,right.

So, in conclusion both models are valid and none is in general better
than the other.  We need however to be aware what model is used in a
particular context.  I think that, for data processing, the center-model
clearly simplifies processing and transformation and that is the reason
why we are using it.

Best,
Stephan




On Sun, 2012-05-13 at 23:24 +0200, Michael Doube wrote: 
> > in Fiji
> > we consider pixel coordinates at a pixel's center which is in perfect
> > agreement when working with microscopic images (particularly confocal
> > scanners where the coordinate really is a point sample
> 
> It's not really - the scanned point is normally moving for the duration
> of sampling, so there is some lateral smear of the sampled space. In
> many confocals the scan speed varies within a row, so the 'point
> samples' aren't  even the same on that level. True point-by-point
> scanning requires the scan mechanism to dwell for the duration of
> measurement, then not measure while the mechanism moves to the next
> dwell point. As far as I'm aware this is not implemented in current
> 'point-scanners' because it would be horribly slow to stop and start the
> scan mirrors millions of times per frame.
> 
> Michael
>