Interesting comparison (300D/40D)

"Wolfgang Weisselberg" wrote:
David J. Littleboy wrote:
"frederick" wrote in message
Charles wrote:
http://www.diwa-labs.com/photoalbum/...=org&id=191587
http://www.diwa-labs.com/photoalbum/...=org&id=191662

To me, it means that the twit who did the test doesn't have a clue as to
what dynamic range means (which is ((max signal)/(noise floor)), or the
log
thereof) or how to test it.

To me, it means it's a case of "meaningless", since the test
is not described well enough for others to repeat it. IOW:
completely worthless.

Presumably, the folks at DIWA blithely plugged their images into DxO and
took what it gave them. So I take it that there's an implicit: follow the
directions that come with DxO and you'll get the same results. Thus I don't
think they are guilty of unrepeatability. I'd bet that if you bought DxO and
played that game with your camera, you'd get similar results.

Simple reality check: Dynamic range must fall by one stop each time the
ISO
doubles.

Must it?
Really?

Please don't snip the important parts. I was quite careful to say:

While this often doesn't hold at low ISOs (due to read noise and
circuit noise), it holds at higher ISOs. And this test doesn't show it.

But, yes. It has to hold. Pretty much by definition.

Is there no possibility that, say, a converter with finite bits
of precision, limits the dynamic range?

Yes. That's what happens at low ISOs. (See figure 5 in the reference below
for the Canon 1DII and 1DIII.)

Or from RAW to JPEG (in camera?) or whatever?

Testing with jpeg is meaningless.

Is there no possibility of the definition of "noise floor"
to be ... wrong,

Exactly! More likely, they don't even think to define it.

I mean, different, in that 'test' due to
noise smoothing? (If you talk about every day performance,
noise supression does play a role!)

Noise reduction entails a loss of resolution. It's equivalent to pixel
binning. So reducing your resolution by a factor of two (1.414x linearly)
should get you one stop of improved DR.

So quoting noise reduced DR values at higher ISOs is seriously stupid.

The 40D seems to deliver ~9 stops between ISO 100 and 1600:
http://www.dpreview.com/reviews/canoneos40d/page20.asp
(which obviously talks about JPEG, since RAW is discussed below).

For the camera to deliver the same "DR" from 100 to 1600, means that they
are changing the definition of the acceptable low end (the noise floor) at
each ISO. That's seriously silly. It's pretty clear that what these tests
are actually measuring is the parameters of the default jpeg conversion.

Again, see figure 5 in the usual reference.

http://www.clarkvision.com/imagedeta...ary/index.html

What I see from Charles' curves is that whatever the limitation
may be, it is probably not the sensor. (But then a camera _is_
more than the sensor.)

This last bit is, of course, important. And part of the camera is the raw
converter; being able to actually get the dynamic range captured in the raw
file into an RGB image file in a manner that's useful is hard and dependent
on the raw converter.

David J. Littleboy
Tokyo, Japan

David J. Littleboy wrote:
"frederick" wrote in message
Charles wrote:

http://www.diwa-labs.com/photoalbum/...=org&id=191587
http://www.diwa-labs.com/photoalbum/...=org&id=191662
So what does this mean to you?

To me, it means that the twit who did the test doesn't have a clue as to
what dynamic range means (which is ((max signal)/(noise floor)), or the log
thereof) or how to test it.

Simple reality check: Dynamic range must fall by one stop each time the ISO
doubles. While this often doesn't hold at low ISOs (due to read noise and
circuit noise), it holds at higher ISOs.

Why?

And this test doesn't show it.

Or you're mistaken. As a rule, I don't support such unsupported
claims as yours. Especially when they have an obvious bias.

--
Ray Fischer

David J. Littleboy wrote:
"Wolfgang Weisselberg" wrote:

http://www.diwa-labs.com/photoalbum/...=org&id=191587
http://www.diwa-labs.com/photoalbum/...=org&id=191662

Presumably, the folks at DIWA blithely plugged their images

JPEGs, I guess ... :-

Or from RAW to JPEG (in camera?) or whatever?

Testing with jpeg is meaningless.

If you (want to) use out-of-camera JPEGs, then testing them is
anything but. If you only want to display the images on a TV
(seriously stupid, I know), you'd best test that setup --- but
never try to generalize from it.

noise smoothing? (If you talk about every day performance,
noise supression does play a role!)

Noise reduction entails a loss of resolution.

However, not all resolution gives detail.

But which does and which does not? That's what a noise reducer
has to guess.

It's equivalent to pixel binning.

Pixel binning can happen before digitisation (and quantisation),
thus incurring the read noise only once --- which noise
reduction working on the digitized image data cannot.

Noise reduction can act differently on local spatial frequencies
(and a lot of other things) which pixel binning cannot.

I content that noise reduction and pixel binning are somewhat
similar, but not equivalent.


The 40D seems to deliver ~9 stops between ISO 100 and 1600:
http://www.dpreview.com/reviews/canoneos40d/page20.asp
(which obviously talks about JPEG, since RAW is discussed below).

For the camera to deliver the same "DR" from 100 to 1600, means that they
are changing the definition of the acceptable low end (the noise floor) at
each ISO. That's seriously silly. It's pretty clear that what these tests
are actually measuring is the parameters of the default jpeg conversion.

Yep. Which is useful to know for customers who want to do
(mosty/only) that.

Again, see figure 5 in the usual reference.

http://www.clarkvision.com/imagedeta...ary/index.html

Yep --- after ISO 1000+ the halving is evident in the 1D.
Fancy that, above ISO 1000 even the diva-curves show
approximate halfing: 8.5, 7.8, 7.1; 10.1, 9,2

-Wolfgang

"Ray Fischer" wrote:
David J. Littleboy wrote:
"frederick" wrote in message
Charles wrote:

http://www.diwa-labs.com/photoalbum/...=org&id=191587
http://www.diwa-labs.com/photoalbum/...=org&id=191662
So what does this mean to you?

To me, it means that the twit who did the test doesn't have a clue as to
what dynamic range means (which is ((max signal)/(noise floor)), or the
log
thereof) or how to test it.

Simple reality check: Dynamic range must fall by one stop each time the
ISO
doubles. While this often doesn't hold at low ISOs (due to read noise and
circuit noise), it holds at higher ISOs.

Why?

Because of the definitions of "dynamic range" and "stop" and "ISO".

And this test doesn't show it.

Or you're mistaken. As a rule, I don't support such unsupported
claims as yours. Especially when they have an obvious bias.

Yes. I'm biased towards technically correct definitions, since anything that
isn't technically correct is BS...

David J. Littleboy
Tokyo, Japan

"Wolfgang Weisselberg" wrote:
David J. Littleboy wrote:
"Wolfgang Weisselberg" wrote:

http://www.diwa-labs.com/photoalbum/...=org&id=191587
http://www.diwa-labs.com/photoalbum/...=org&id=191662

Presumably, the folks at DIWA blithely plugged their images

JPEGs, I guess ... :-

Or from RAW to JPEG (in camera?) or whatever?

Testing with jpeg is meaningless.

If you (want to) use out-of-camera JPEGs, then testing them is
anything but.

But if you are interested in (more) dynamic range, then you won't be
shooting jpeg. Jpeg has to throw away at least two stops of dynamic range
(at ISO 100) to produce an image with enough contrast to be at all
attractive.

noise smoothing? (If you talk about every day performance,
noise supression does play a role!)

Noise reduction entails a loss of resolution.

However, not all resolution gives detail.

I don't see what you are getting at???

But which does and which does not? That's what a noise reducer
has to guess.

It's equivalent to pixel binning.

Pixel binning can happen before digitisation (and quantisation),
thus incurring the read noise only once --- which noise
reduction working on the digitized image data cannot.

Noise reduction can act differently on local spatial frequencies
(and a lot of other things) which pixel binning cannot.

Oh. Now I see. You're saying that noise reduction will be applied
intelligently so that places without detail will get smoothed, and places
with detail won't get mucked with as much.

Maybe. But the pre-noise reduction signal has a certain DR, and claiming
that you get something for nothing with "intelligent NR" strikes me as
suspicious in the extreme.

Anyway, "DR" is a property of individual pixels, and claiming that it's
changed by certain post-processing is odd. If noise reduction on an image
produced with "8-stop DR pixels" produces a "10-stop DR total image", why
doesn't applying the same processing to "10-stop DR pixels" produce a
"12-stop DR total image".

I content that noise reduction and pixel binning are somewhat
similar, but not equivalent.

That could be. But the bottom line is that a technical parameter like DR
should be measured on an un-postprocessed signal, since presumably
the "intelligent NR" will do different things to different images.

The 40D seems to deliver ~9 stops between ISO 100 and 1600:
http://www.dpreview.com/reviews/canoneos40d/page20.asp
(which obviously talks about JPEG, since RAW is discussed below).

For the camera to deliver the same "DR" from 100 to 1600, means that they
are changing the definition of the acceptable low end (the noise floor)
at
each ISO. That's seriously silly. It's pretty clear that what these tests
are actually measuring is the parameters of the default jpeg conversion.

Yep. Which is useful to know for customers who want to do
(mosty/only) that.

Again, see figure 5 in the usual reference.

http://www.clarkvision.com/imagedeta...ary/index.html

Yep --- after ISO 1000+ the halving is evident in the 1D.

The halving starts at ISO 400; the 800 DR is essentially a whole stop below
ISO 400. The other cameras are straight lines throughout the range. And both
the 40D and D300 are smaller-pixel cameras.

Fancy that, above ISO 1000 even the diva-curves show
approximate halfing: 8.5, 7.8, 7.1; 10.1, 9,2

Oops. You're right: there is the expected falloff at the far end of these
graphs.

Still, they have the D300 being pretty flat out to 1600 which is definitely
Journal of Irreproducible Results territory, as is claiming 10 stops of DR
for the 40D at ISO 1600.

Again, these aren't the fat-pixel D3/5D/1DII cameras, they're D200 and
smaller pixels.

David J. Littleboy
Tokyo, Japan

David J. Littleboy wrote:
Anyway, "DR" is a property of individual pixels, ...
... DR should be measured on an un-postprocessed signal

So is DR a property of a single pixel (presumably within the sensor, or the
signal from that sensor) or of the signal from, presumably, multiple pixels?
I can imagine a definition that refers to either the capability of the
device, or of the device as it may be reduced by following processing, or of
an actual signal obtained via that route on some specific occasion. Which
would surely be in order of decreasing DR since an instance cannot exceed a
capability. Which you mean doesn't matter to me but you seem to be sure what
you mean and I am not...

Oh, yes, and is DR referred back to light intensity?

Mike.
--
If reply address = connectfee, add an r because it is free not fee.

David J. Littleboy wrote:
"Wolfgang Weisselberg" wrote:
David J. Littleboy wrote:
"Wolfgang Weisselberg" wrote:

Testing with jpeg is meaningless.

If you (want to) use out-of-camera JPEGs, then testing them is
anything but.

But if you are interested in (more) dynamic range,

Most people aren't. Most people use P&S cameras and many of
those wouldn't know exposure from aperture if it bit them in
the nose. :-)
And enough people are perfectly fine with out-of-camera JPEGs.

Noise reduction entails a loss of resolution.

However, not all resolution gives detail.

I don't see what you are getting at???

How much resolution do you need for a nicely blurred, out of
focus, background? When do you stop getting more detail from
more resolution? Could you trade that extra resolution for less
noise without harming the image?

Oh. Now I see. You're saying that noise reduction will be applied
intelligently so that places without detail will get smoothed, and places
with detail won't get mucked with as much.

That's the idea, AKA "the easy part".

Maybe. But the pre-noise reduction signal has a certain DR, and claiming
that you get something for nothing with "intelligent NR" strikes me as
suspicious in the extreme.

You don't get "something for nothing".
You trade resolution you (hopefully) don't need for less noise,
thus lowering the noise floor (at least in low detail areas,
like ... evenly lit gray or black parts designed to measure noise).

Anyway, "DR" is a property of individual pixels,

While I am quite sure no 'regular' camera does this today, I can
imagine one that, guided by it's high resolution matrix metering,
- applies analog gain differently per matrix area, and/or
- uses an electronic shutter and different exposure times per
matrix area.

This would increase overall DR, unless you have the brightest
and darkest parts too close for the metering to differenciate.
Yet the per-pixel-DR would be untouched.

and claiming that it's
changed by certain post-processing is odd. If noise reduction on an image
produced with "8-stop DR pixels" produces a "10-stop DR total image", why
doesn't applying the same processing to "10-stop DR pixels" produce a
"12-stop DR total image".

Assuming "intelligent NR" behaves easily predictable and/or
linearly is assuming a naive "intelligent" NR.
However, good NR can try to reduce the relevant[1] noise floor.

-Wolfgang

[1] relevant in the same sense that lossy image compressions
like JPEG remove irrelevant visual data and lossy audio
compressions (e.g. Ogg Vorbis) tries to keep only the
relevant data.

David J. Littleboy wrote:
"Ray Fischer" wrote:
David J. Littleboy wrote:
"frederick" wrote in message
Charles wrote:

http://www.diwa-labs.com/photoalbum/...=org&id=191587
http://www.diwa-labs.com/photoalbum/...=org&id=191662
So what does this mean to you?

To me, it means that the twit who did the test doesn't have a clue as to
what dynamic range means (which is ((max signal)/(noise floor)), or the
log
thereof) or how to test it.

Simple reality check: Dynamic range must fall by one stop each time the
ISO
doubles. While this often doesn't hold at low ISOs (due to read noise and
circuit noise), it holds at higher ISOs.

Why?

Because of the definitions of "dynamic range" and "stop" and "ISO".

Spare me the bull****. If you don't know then just say so.

And this test doesn't show it.

Or you're mistaken. As a rule, I don't support such unsupported
claims as yours. Especially when they have an obvious bias.

Yes. I'm biased towards technically correct definitions,

You're biased towards bull****.

--
Ray Fischer

David J. Littleboy wrote:
Ray Fischer wrote:
David J. Littleboy wrote:

Simple reality check: Dynamic range must fall by one stop each
time the ISO doubles. While this often doesn't hold at low ISOs
(due to read noise and circuit noise), it holds at higher ISOs.

Why?

Because of the definitions of "dynamic range" and "stop" and "ISO".

It would help me a lot to understand what you mean if those definitions were
stated here. Please.

"Wilba" wrote:
David J. Littleboy wrote:
Ray Fischer wrote:
David J. Littleboy wrote:

Simple reality check: Dynamic range must fall by one stop each
time the ISO doubles. While this often doesn't hold at low ISOs
(due to read noise and circuit noise), it holds at higher ISOs.

Why?

Because of the definitions of "dynamic range" and "stop" and "ISO".

It would help me a lot to understand what you mean if those definitions
were stated here. Please.

For our purposes, dynamic range is log2((max signal)/(noise floor)).

There are a lot of definitions of dynamic range depending on the purpose at
hand, but they all are the ratio of the (max signal) to the (noise floor)
expressed in a manner that's useful to the problem at hand.

In engineering, the concept is well understood. A quick google came up with
these and lots more.

http://www.its.bldrdoc.gov/fs-1037/dir-013/_1853.htm
http://whatis.techtarget.com/definit...13623,00.html#

But when photography appears, people get confused. Here's a page full of
links to various levels of confusion.

http://en.mimi.hu/photography/dynamic_range.html

David J. Littleboy
Tokyo, Japan