FF sensors: is 80MP needed?

Scott W wrote:
Roger N. Clark (change username to rnclark) wrote:

Raphael Bustin wrote:


Unless I'm wrong, R-L requires characterizing lens distortion.

No, it is any distortion. You can even do motion blur.

But really now... are we going to do that over a continuum of
focal lengths, focus distances, and apertures? Will the
characterization be constant throughout the image circle?

Will we repeat this process for each lens in the kit, and
record all taking parameters, etc.?

I mean, seriously... how well can a lens (not to mention
a zoom lens) be characterized?

Roger's fox-eye comparisons haven't sold me. In particular,
Roger's choice of USM parameters is strange. For that
particular image sample -- after careful upsampling -- I'd
have used a much smaller radius and much higher
amount. (Eg. radius 1.0 and amount = 400%).

You are welcome to download the image and try. Many have.
Only Bill Hilton has equaled the R-L results using
edge detection masks plus unsharp mask (I gotta add that
result to the web page sometime).


I have done this one before, my USM comes I believe very close to
yours.
http://www.pbase.com/konascott/image/54355011/original
But what I do is to convert the image into a linear space before
applying the USM
and then back to sRGB after. For some odd reason PS is assuming that
the pixels are linear when doing USM, which is why you get many of the
odd artifacts, light really bright pixels.

Scott

Hi Scott,
At first look, your image does appear pretty close.
I put your image into photoshop and diffed it with the R-L
image (Figure 6 from the web page) and found a strange
odd-even pattern. Then looking at the original two
images (yours and Figure 6), the up sampling artifacts
are quite apparent, and your image has squarish pixels
double the size of the Figure 6 pixels. Put the two
images in photoshop and look at the eye at 300%

What up-sampling algorithm did you use?

Roger

Raphael Bustin wrote:

On Mon, 16 Oct 2006 22:53:07 -0700, "Roger N. Clark (change username
to rnclark)" wrote:



You are welcome to download the image and try. Many have.
Only Bill Hilton has equaled the R-L results using
edge detection masks plus unsharp mask (I gotta add that
result to the web page sometime).



I did exactly that, Roger. Upon downloading the
image, I upsampled in four or five steps, and then
sharpened with Amt=400-500, Radius=0.8-1.2.

I find my parameters much less garish, and much
more subtle than your USM sample. Yours does
show more "apparent" sharpness, but also more
artifacts. Bottom line: if faced with the same
upsampling reqiurement, I wouldn't dare sharpen
to the extent that you did with USM.

In both your USM sample and in your RL sample,
there is visible staircasing on some of the fox's
whiskers and facial hairs.

Rafe,
Yes, I agree with you on the staircase problems.
But those are not due to R-L, those are due
to the up-sampling algorithm. The up-sampling
was a simple cubic spline in Photoshop. I have
not been happy with cubic cpline but haven't seen
anything better. Have you?

You must be aware that these are the two most
contentious and subjective issues in digital image
processing. RL gives a sharp fox, but it's still
not a real fox.

Yes, I agree. But the fox wasn't "real" before any processing
either. The camera optics smeared the detail, and the
raw converter interpolated RGB pixels to make the
RGB tif, producing other artifacts along the way.

Roger

Roger N. Clark (change username to rnclark) wrote:
Scott W wrote:
Roger N. Clark (change username to rnclark) wrote:

Raphael Bustin wrote:


Unless I'm wrong, R-L requires characterizing lens distortion.

No, it is any distortion. You can even do motion blur.

But really now... are we going to do that over a continuum of
focal lengths, focus distances, and apertures? Will the
characterization be constant throughout the image circle?

Will we repeat this process for each lens in the kit, and
record all taking parameters, etc.?

I mean, seriously... how well can a lens (not to mention
a zoom lens) be characterized?

Roger's fox-eye comparisons haven't sold me. In particular,
Roger's choice of USM parameters is strange. For that
particular image sample -- after careful upsampling -- I'd
have used a much smaller radius and much higher
amount. (Eg. radius 1.0 and amount = 400%).

You are welcome to download the image and try. Many have.
Only Bill Hilton has equaled the R-L results using
edge detection masks plus unsharp mask (I gotta add that
result to the web page sometime).


I have done this one before, my USM comes I believe very close to
yours.
http://www.pbase.com/konascott/image/54355011/original
But what I do is to convert the image into a linear space before
applying the USM
and then back to sRGB after. For some odd reason PS is assuming that
the pixels are linear when doing USM, which is why you get many of the
odd artifacts, light really bright pixels.

Scott

Hi Scott,
At first look, your image does appear pretty close.
I put your image into photoshop and diffed it with the R-L
image (Figure 6 from the web page) and found a strange
odd-even pattern. Then looking at the original two
images (yours and Figure 6), the up sampling artifacts
are quite apparent, and your image has squarish pixels
double the size of the Figure 6 pixels. Put the two
images in photoshop and look at the eye at 300%

What up-sampling algorithm did you use?

It has been a while since I did that but I believe I first translated
to linear space used good old bicubic for up sampling then used USM and
then translated back to sRGB. USM works like crap when working on sRGB
but does ok when working in a linear space. A good test of this is
using USM on uniform noise, when doing it on an sRGB image the over all
brightness increases, clearly this should not be the case. If first
convert to linear space and then use USM and then convert back the over
all brightness does not change. This is a really silly bug in PS and
it would be easy for them to fix it. The halos that come from using
USM are also not nearly as bad when using it in linear space.

Scott

Scott W wrote:
David J. Littleboy wrote:

This is very odd. It sounds as though you don't, at the gut level, accept
that diffraction actually destroys detail.

I when round and round on this with Roger and just gave up in the end.
I think Roger is believing he is seeing detail in the grass part of his
photos that is in
fact grain, this is pretty clear since the diffraction spot is lager
then some of the detail in the grass.

Scott

Scott,
The diffraction spot in an image CAN appear smaller than
the full diffraction spot. A diffraction spot has
a narrow peak. With high contrast film, like Velvia,
you can lose the bottom of the diffraction pattern
leaving only the peak. You can try this yourself:
generate a diffraction pattern, then slice off the lower end.
The higher you slice, the smaller the remaining spot
appears.

While diffraction limits detail, the diffraction limit (Dawes,
or 0% MTF) does NOT mean no detail. Example: two point sources,
or two lines close together have a broadened combined diffraction
pattern. But the fact that the pattern IS broadened
is information from which you can recover information
about the sources. The separation of stars in
telescopes can be determined when they are closer than
the Dawes limit (the 0% MTF point).

Second in the equation of image detail is
sampling and the phase of the sampling
by the scanner pixels. If you choose pixels spaced at just above
the Dawes limit, in theory you get all the information,
but only if the image detail falls perfectly on the scanner pixels.
Statistically it does not. A good demonstration of that
is detection of an edge that is not aligned to the pixel
grid. The higher you sample that edge, the closer you
can come to recording the information recorded by the
optical system, including sampling above the 0% MTF.
0% MTF means you can't resolve a grid of lines
at that frequency; it does not mean you can't
infer/compute or even visually see
two closely spaced lines or points.

Phase effects are illustrated in the grid at
(first figure on the page):
http://www.clarkvision.com/imagedetail/sampling1.html
Below 3x Nyquist sampling phasing errors cause image
artifacts. The difference here and in
scanned film, or digital cameras, is the sampling is
finite, and that changes the problem from the Nyquist
theorem.

So, then in the grass field image on the above web page,
it you stand back 15 feet or more, and tell me which
image looks the sharpest. I know for a fact, from
being at the scene, that the grass field at the top of
the image has long blades of grass sticking up. The
curvy-linear strands of bright posts are those grass
blades below the diffraction limit. The image is a
combined effect of grain clumps limiting resolution, and
diffraction spots. So then the question is what
ppi is needed to record that detail. There are 4
levels shown on the figure. Which one do you think
is adequate?

Roger

David J. Littleboy wrote:

"Scott W" wrote in message
oups.com...

David J. Littleboy wrote:

This is very odd. It sounds as though you don't, at the gut level, accept
that diffraction actually destroys detail.

I went round and round on this with Roger and just gave up in the end.
I think Roger is believing he is seeing detail in the grass part of his
photos that is in
fact grain, this is pretty clear since the diffraction spot is lager
then some of the detail in the grass.


Yep. I did too; I couldn't make any sense out of his examples that he
thought were showing more detail; I'm just being slower in figuring that
out.

I suspect that the problem here (especially with the R-L stuff) is that he
thinks about images as though they were star pattern. But it's a lot easier
to "resolution enhance" star patterns since to a very large degree you know
what you are looking for. But since you don't have anywhere near that good
an idea what pictorial images are, the resolution enhancement game doesn't
apply. (This is naively expressed, and the math probably claims that each
point in a pictorial image acts like a low brightness star; but I suspect
that you have to handle a small number of (extremely intense) true points
very differently from smoothly changing gradients with an occasional
low-contrast edge.)

That's what R-L does. It is an iterative process.
Perhaps try reading one of the technical papers:

http://www.stsci.edu/stsci/meetings/...er_damped.html

This page smears images then reconstructs them:
http://www.adass.org/adass/proceedin.../prukschm.html

Algorithms like R-L have been in development for decades
with many scientific papers written on them. The theory
is well developed and they work well in practice.
But like any algorithm, it needs good data and should
not be abused.

Roger

[A complimentary Cc of this posting was sent to
Roger N. Clark (change username to rnclark)
], who wrote in article :
Digital has a much higher signal-to-noise ratio

Such a blank statement is definitely wrong; you mean the same
exposition, and it is not applicable to my comparison. E.g, with the
current technology digital will give much worse noise than film if the
exposition of digital is 20x smaller than of film ;-). And note that
this is quite probably holds when you compare LF film to FF digital
(but I do not remember the details of the exposition you used...) -
count number of photons per "pixel" (I mean square with size
determined by the MTF curve).

By ANY measure that is even close to similar between
digital and film, digital wins.

In what I investigated, I studied the case when digital will be able
to win EVEN IF is the situation NOT IN ANY WAY close to film. E.g.,
4x5in film with f/64 vs FF sensor with f/1.

Film has practically no noise when correctly exposed with aperture
f/32 - or much smaller f-number.

Well, that is simply not true. Film has grain and grain
clumps. That translates to irregularities people see in
images, of variations measured by a densitometer.
And it has nothing to do with f/stop.

Just the opposite. E.g., the variations measured by a densitometer
will be very different depending on the area over which it averages
the density. And with f/64 the area of e.g., diffraction circle is so
large that the variations will be negligible.


IMO, this remains to be proven yet (though I suspect that the proof
*will* support your claim - I just *have not seen* it).

e.g.:
Dynamic Range and Transfer Functions of Digital Images
and Comparison to Film
http://www.clarkvision.com/imagedetail/dynamicrange2
See Figures 5 and 8. I n figure 8, not how noisy the film
characteristic curve is compared to the digital (blue dots).

As I discussed below, this investigation has very little to do with
what most people will call dynamic range. A measurement of
high-spacial-frequency-Noise is not a measurement of dynamic range.

It doesn't matter how you define it, as long as you define
it equally for all media. Digital wins in all cases,
because if it wins on a pixel basis, it wins when you
average larger areas.

Remember claims "one needs 16MP sensor to get similar performance to
the film"? It was very fine - just wrong. Your argument is very
convincing, but still "just an argument"; there is A LOT of
assumptions needed to true for your argument to work (e.g. the same
spectrum of noise).

As I said, until it is ACTUALLY measured, there is very little hope to
establish it "just by theoretizing".

One gets a curve of (thus defined) spacial resolution per degree of
underexposition; *this curve*, IMO, is the measure of dynamic range.

Again: I agree with you that such an investigation will *most
probably* give film much lower mark than digital sensor - but 10x20in
film is practical, and digital is not.

If you have to go to 10x20 inch film to beat a small digital
sensor in dynamic range, it is not worth much. In the time
one took to slide the 10x20 inch plate holder into place,
one could take 2 different exposures providing 15 stops of
dynamic range. Gee you could probably take 4 or 5 exposures,
getting greater than 25 stops. Not that one needs that
most of the time.

While your original post referenced my digital mosaics
web article, I feel you were quite off in the f/ratio
tables you posted. I routinely make 30x40 inch prints
from f/32 to f/64 4x5 Velvia transparencies. I've even
done 40 x 60 inch prints from f/45 4x5s. The detail is quite
impressive.

The details at the particular f/64 shot I investigated correspond to
14MP JPEG. Frankly speaking, I investigated only the crops you
provided on this web page; are they in slightly-out-of-focus area
then, or what?

My own testing, with is documented on my
web pages at http://www.clarkvision.com/imagedetail
puts 4x5 Velvia at 200 megapixel digital equivalent.

I did not investigate 4x5 Velvia. I investigated 4x5 Velvia with f/64
lens. It corresponds to about 15x lower MP count.

Hope this helps,
Ilya

[A complimentary Cc of this posting was NOT [per weedlist] sent to
Ilya Zakharevich
], who wrote in article :
[A complimentary Cc of this posting was NOT [per weedlist] sent to
By ANY measure that is even close to similar between
digital and film, digital wins.

In what I investigated, I studied the case when digital will be able
to win EVEN IF is the situation NOT IN ANY WAY close to film. E.g.,
4x5in film with f/64 vs FF sensor with f/1.
^^^^^^^^

Of course, I meant f/16. (Do I have dust under the "6" key? ;-)

Sorry,
Ilya

[A complimentary Cc of this posting was sent to
frederick
], who wrote in article 1161034115.40328@ftpsrv1:
I do not agree with you. Read: "on the plus side"; obviously, he
thinks that with smaller formfactor he will have larger depth of
field. (This is only true if he would use the same f-stop; but when
you convert to a more open f-stop needed to get the same diffraction
[measured in the subject space, i.e., in angular units], one gets
exactly the same DoF...)

Yes - but he also gets increased DOF at the same f-stop , same FOV, and
same shutter speed.

Still, I can't agree with you. The only reason why DOF increases is
that in-focus areas contain about 2.5x less linear detail, thus you
may go further from the focus plane until the defocus circle of
confusion reaches this size.

I can hardly call it "same something" - although, of course, on paper
you are 100% correct - you just did not mention how much blurrier the
image is. ;-)

Ultimately you are of course correct - but it may mean that you can't
take a photo with even a moving snail in the frame ;-)

Now will start to violently agree on what the other guy saiz... ;-)

Yours,
Ilya

"Roger N. Clark (change username to rnclark)" writes:
Yes, I agree with you on the staircase problems. But those are not
due to R-L, those are due to the up-sampling algorithm. The
up-sampling was a simple cubic spline in Photoshop. I have not been
happy with cubic cpline but haven't seen anything better. Have you?

I thought that fancier upconverters used some interpolation involving
the sinc function.

Scott W wrote:
Roger N. Clark (change username to rnclark) wrote:
Scott W wrote:
Roger N. Clark (change username to rnclark) wrote:

Raphael Bustin wrote:


Unless I'm wrong, R-L requires characterizing lens distortion.
No, it is any distortion. You can even do motion blur.

But really now... are we going to do that over a continuum of
focal lengths, focus distances, and apertures? Will the
characterization be constant throughout the image circle?

Will we repeat this process for each lens in the kit, and
record all taking parameters, etc.?

I mean, seriously... how well can a lens (not to mention
a zoom lens) be characterized?

Roger's fox-eye comparisons haven't sold me. In particular,
Roger's choice of USM parameters is strange. For that
particular image sample -- after careful upsampling -- I'd
have used a much smaller radius and much higher
amount. (Eg. radius 1.0 and amount = 400%).
You are welcome to download the image and try. Many have.
Only Bill Hilton has equaled the R-L results using
edge detection masks plus unsharp mask (I gotta add that
result to the web page sometime).

I have done this one before, my USM comes I believe very close to
yours.
http://www.pbase.com/konascott/image/54355011/original
But what I do is to convert the image into a linear space before
applying the USM
and then back to sRGB after. For some odd reason PS is assuming that
the pixels are linear when doing USM, which is why you get many of the
odd artifacts, light really bright pixels.

Scott

Hi Scott,
At first look, your image does appear pretty close.
I put your image into photoshop and diffed it with the R-L
image (Figure 6 from the web page) and found a strange
odd-even pattern. Then looking at the original two
images (yours and Figure 6), the up sampling artifacts
are quite apparent, and your image has squarish pixels
double the size of the Figure 6 pixels. Put the two
images in photoshop and look at the eye at 300%

What up-sampling algorithm did you use?

It has been a while since I did that but I believe I first translated
to linear space used good old bicubic for up sampling then used USM and
then translated back to sRGB. USM works like crap when working on sRGB
but does ok when working in a linear space. A good test of this is
using USM on uniform noise, when doing it on an sRGB image the over all
brightness increases, clearly this should not be the case. If first
convert to linear space and then use USM and then convert back the over
all brightness does not change. This is a really silly bug in PS and
it would be easy for them to fix it. The halos that come from using
USM are also not nearly as bad when using it in linear space.

Scott

Have you (or Roger) looked at this:
http://www.general-cathexis.com/