| Re: fitting after rebinning [message #57343] |
Thu, 06 December 2007 17:14 |
Greg Hennessy
Messages: 45
Registered: November 2005
|
Member |
|
|
On 2007-12-07, biophys <biophys@gmail.com> wrote:
> Well I do this kinda
> fitting almost everyday with CCD images. If you can imagine each data
> point of the image array as the readout of a CCD array, you will
> understand this almost immediately.
Well, I should have understood it immediately.
> To proper scale the results before and after rebin you just have
> to get rid of the half pixel "offset" and do the scale and add back
> the half pixel back after scaling. e.g.
> 255.5+0.5=(31.5+0.5)*(512/64)=(7.5+.5)*(512/16)
I had been trying to subtract the half pixel, and my math wasn't
working, adding it would have given me the math you showed here.
I really hate dealing with those half pixel offsets.
|
|
|
|
| Re: fitting after rebinning [message #57344 is a reply to message #57343] |
Thu, 06 December 2007 16:17  |
biophys
Messages: 68
Registered: July 2004
|
Member |
|
|
Yes. Using /sample with compression is "morally" wrong in most cases.
It makes sense only when you have a need for speed and don't care
about the details. The fitting you are doing gives you absolutely the
right answer. But how to understand the results. Well I do this kinda
fitting almost everyday with CCD images. If you can imagine each data
point of the image array as the readout of a CCD array, you will
understand this almost immediately. In other words, the data
represents the readout from the center of a pixel. Now psf_gaussian is
on the other hand centered on the corner of four neighboring pixels
which gives you always the .5 pixel offset. Let's take a look:
IDL>a=psf_gaussian(npix=512,fwhm=90)
IDL>print, a[254:257,254:257]
0.998461 0.999145 0.999145 0.998461
0.999145 0.999829 0.999829 0.999145
0.999145 0.999829 0.999829 0.999145
0.998461 0.999145 0.999145 0.998461
IDL>b=rebin(a,64,64)
IDL>print, b[30:33,30:33]
0.903193 0.943497 0.943497 0.903193
0.943497 0.985599 0.985599 0.943497
0.943497 0.985599 0.985599 0.943497
0.903193 0.943497 0.943497 0.903193
IDL> c=rebin(a,64,64,/sample)
IDL> print,c[30:33,30:33]
0.848341 0.903490 0.920975 0.898555
0.903490 0.962224 0.980846 0.956968
0.920975 0.980846 0.999829 0.975489
0.898555 0.956968 0.975489 0.951742
IDL> t=gauss2dfit(a,p1,/tilt)
Compiled module: GAUSS2DFIT.
Compiled module: GAUSSFIT.
Compiled module: POLY_FIT.
Compiled module: CURVEFIT.
Program caused arithmetic error: Floating divide by 0
Program caused arithmetic error: Floating underflow
IDL> print,p1
2.19443e-008 0.999914 38.2195 38.2195 255.500
255.500 0.000000
IDL>t=gauss2dfit(b,p2,/tilt)
CURVEFIT: Failed to converge- CHISQ increasing without bound.
Program caused arithmetic error: Floating divide by 0
Program caused arithmetic error: Floating illegal operand
IDL> print,p2
1.54614e-007 0.990991 4.78601 4.78601 31.5000
31.5000 0.000000
As you can see after rebin(w/o /sample), the center is still located
at the corner of a pixel. It makes sense that if you think pixel No 0
is located at [-0.5,0.5]x[-0.5x0.5] and so on so that [255.5,255.5] or
[31.5,31.5] is exactly what you would expect where the center should
be. To proper scale the results before and after rebin you just have
to get rid of the half pixel "offset" and do the scale and add back
the half pixel back after scaling. e.g.
255.5+0.5=(31.5+0.5)*(512/64)=(7.5+.5)*(512/16)
Cheers,
BP
On Dec 6, 1:53 pm, Greg Hennessy <greg.henne...@localhost.localdomain>
wrote:
>>> The offsets go away. Hmm, I need to think on this.
>
>> I would consider your assumptions on what happens when
>> you rebin. :-)
>
> Yea. If I call A the big image, B a rebin image not using sample, and
> C the rebin,/sample image, when I plotimage the three A and B look
> the most alike, with C having an offset, but the fitted parameters
> of A and C seem more alike. Like i said, I need to think.
|
|
|
|
| Re: fitting after rebinning [message #57350 is a reply to message #57344] |
Thu, 06 December 2007 14:13 |
Greg Hennessy
Messages: 45
Registered: November 2005
|
Member |
|
|
On 2007-12-06, David Fanning <david@dfanning.com> wrote:
> Wasn't it Yoda who said "Believe your head, Luke, not
> your eyes!"
Probably Obi Wan said that. If it were Yoda it would have been more
like "Believe head, Luke, your eyes not".
Or that line from Groucho, "Who are you going to believe, me or your
lying eyes?"
|
|
|
|
| Re: fitting after rebinning [message #57352 is a reply to message #57350] |
Thu, 06 December 2007 13:58 |
David Fanning
Messages: 11724
Registered: August 2001
|
Senior Member |
|
|
Greg Hennessy writes:
>>> The offsets go away. Hmm, I need to think on this.
>>
>> I would consider your assumptions on what happens when
>> you rebin. :-)
>
> Yea. If I call A the big image, B a rebin image not using sample, and
> C the rebin,/sample image, when I plotimage the three A and B look
> the most alike, with C having an offset, but the fitted parameters
> of A and C seem more alike. Like i said, I need to think.
Wasn't it Yoda who said "Believe your head, Luke, not
your eyes!"
Cheers,
David
--
David Fanning, Ph.D.
Fanning Software Consulting, Inc.
Coyote's Guide to IDL Programming: http://www.dfanning.com/
|
|
|
|
| Re: fitting after rebinning [message #57353 is a reply to message #57352] |
Thu, 06 December 2007 13:53 |
Greg Hennessy
Messages: 45
Registered: November 2005
|
Member |
|
|
>> The offsets go away. Hmm, I need to think on this.
>
> I would consider your assumptions on what happens when
> you rebin. :-)
Yea. If I call A the big image, B a rebin image not using sample, and
C the rebin,/sample image, when I plotimage the three A and B look
the most alike, with C having an offset, but the fitted parameters
of A and C seem more alike. Like i said, I need to think.
|
|
|
|
|
|
| Re: fitting after rebinning [message #57356 is a reply to message #57355] |
Thu, 06 December 2007 13:41 |
Greg Hennessy
Messages: 45
Registered: November 2005
|
Member |
|
|
On 2007-12-06, Brian Larsen <balarsen@gmail.com> wrote:
> After running your code I get a curve fit error...
Well, I had cheated. I had origionally used mpfit2dpeak instead of
gauss2dfit, but then I thought that someone might assert that Craig's
routine had an error, and I'm sure its not his fitting routine that
has an issue, so I swapped out Craig's routine for the faster and
possibly less accurate gauss2dfit. The two routines give the same
answer, even with the convergence message.
|
|
|
|
|
|
| Re: fitting after rebinning [message #57358 is a reply to message #57357] |
Thu, 06 December 2007 13:21 |
Brian Larsen
Messages: 270
Registered: June 2006
|
Senior Member |
|
|
After running your code I get a curve fit error...
IDL> t1=gauss2dfit(a,p1,/tilt)
% Program caused arithmetic error: Floating divide by 0
% Program caused arithmetic error: Floating underflow
IDL> t2=gauss2dfit(b,p2,/tilt)
% CURVEFIT: Failed to converge- CHISQ increasing without bound.
% Program caused arithmetic error: Floating divide by 0
% Program caused arithmetic error: Floating illegal operand
IDL> t3=gauss2dfit(c,p3,/tilt)
% CURVEFIT: Failed to converge- CHISQ increasing without bound.
% Program caused arithmetic error: Floating divide by 0
% Program caused arithmetic error: Floating illegal operand
Which makes me wonder why it gave an answer at all... if it fails to
converge doesn't that mean that the answer is no good for t2 and t3?
Cheers,
Brian
------------------------------------------------------------ --------------
Brian Larsen
Boston University
Center for Space Physics
|
|
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
