D3X vs 5D2 - myth and reality

Started Jan 26, 2009 | Discussions
JerryCanon10D Regular Member • Posts: 428
Re: and..

jeff thigpen wrote:

Nicely said Jeff!
I'll now slip back to the Canon forums ... until I sell my 1DmkIII & buy a D3x.

ejmartin Veteran Member • Posts: 6,274
Re: Measurements

Do you compensate for any software multiplier that Nikon applies to the R channel? The D300 had the R channel stretched by about 1.25 which is 1/3 stop. Without compensating for this, one would underestimate the G channel DR by 1/3 stop using your method.
--
emil
--

http://theory.uchicago.edu/~ejm/pix/20d/

OP GaborSch Veteran Member • Posts: 7,203
Scaling

ejmartin wrote:

Do you compensate for any software multiplier that Nikon applies to
the R channel? The D300 had the R channel stretched by about 1.25
which is 1/3 stop. Without compensating for this, one would
underestimate the G channel DR by 1/3 stop using your method.

1. there is no need to compensate for that,

b. notwithstanding the above, this does not impede the DR calculation.

DSPographer Senior Member • Posts: 2,466
Re: Measurements

Gabor,

Thanks for the explanation of your method. I still have a question about the black patch levels: How does the black patch level compare to the masked pixel level? If it is one masked pixel standard deviation above the masked pixel level then even with uniform illumination the measured black patch standard deviation has been contaminated by photon shot noise.

GaborSch wrote:

[snip]
As we are talking about dark patches ( minimum 7 EV from
saturation), AND about roughly equal illumination between the two
cameras' images at the same ISOs, the photon noise is negligable.
[snip]

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
bobn2
bobn2 Forum Pro • Posts: 61,439
Re: Scaling

GaborSch wrote:

ejmartin wrote:

Do you compensate for any software multiplier that Nikon applies to
the R channel? The D300 had the R channel stretched by about 1.25
which is 1/3 stop. Without compensating for this, one would
underestimate the G channel DR by 1/3 stop using your method.

1. there is no need to compensate for that,

b. notwithstanding the above, this does not impede the DR calculation.

Hang on, Gabor. You said you didn't make a DR calculation. If you did, how do you make it from your measurements?
--
Bob

DSPographer Senior Member • Posts: 2,466
Gabor's channel independence post below shows no stretching

If one channel of the D3X had a different analog or digital (stretching) gain than the other channels then it would represent a different electron count and thus would have a different shot noise standard deviation. To within the square of the precision Gabor shows the standard deviations are equal then the gains must also be equal.

 DSPographer's gear list:DSPographer's gear list
Canon EOS 5D Mark II Canon EOS M5 Canon EF 24mm f/2.8 Canon EF 50mm f/1.8 II Canon EF 200mm f/2.8L II USM +4 more
OP GaborSch Veteran Member • Posts: 7,203
Re: Measurements

bobn2 wrote:

The difference has been observed and demostrated by Emil and others,
so at the moment it would seem that your measurement is the outlier

As I am measuring this instead of speculating about it, I have no problem whatsoever.

Read noise, sure. As for DR (more on that later) this depends on full
scale count (not necessarily the same as saturation) which can be
different from channel to channel. What this means is that a camera's
DR does indeed depend on the actual illuminant - hence Iliah Borg's
magenta filter trick

Sorry, you are mixing up two different issues.

1. Calculating the actual intensity, relative to the saturation point: you need to know, when saturation occurs. I do know that (though sometimes some camera copies contradict that). Anyway, don't mix up the binary numerical range with teh saturation levels.

2. The composition of the illuminant. I have a magenta filter, which I use (extremely seldom) to increase the dynamic range of the camera , NOT of the sensor .

Transputed?

Transposed?

Let's put it another way, how do you discover the 'saturation points'?

I am looking for raw files with saturated areas.

I have not calculated the DR at all.

I find this difficult to square with your statement 'The result
proves, that the claim "two stops higher DR than the 5D2" is
ridiculous , like some other claims. In fact, the DR of the D3X is
max. 0.5 EV greater than that of the 5D2 at ISO 100, and at ISO 400
the 5D2 is already better.'

Back to my statement: the DR depends on the ISO, on the subject and on the actual illumination, AND on your acceptance of noise.

I am not saying how high the DR of this or that camera is. I am saying, that this or that camera creates a certain level of noise at a certail level of illumination (measured on the raw channel) . The difference between the D3X and the 5D2 is FAR from the two stops, even from one stop.

That's an SNR figure, not a DR. I've never been able to convince
myself that DxO's 18% grey noise value is a very useful metric

Not have I. I don't give a fig for how DxO measures the noise.

True, but it's always proportional to the ratio between full-scale
and read noise. All that happens is you might set your 'acceptable'
lower limit some amount above the noise floor

???

On many cameras 7EV from saturation is 6EV or more above the read
noise floor.
Shot noise will be about 3 stops greater than read noise at that level.
(Edit: I should have said 5 stops above read noise floor and shot
noise 2.5 stops above read noise)

Consequence?

The effect is slight, but at the same exposure the shot noise in 900
pixels of 5DII will be lower than 900 pixels of D3x

Sorry, this is far out of the range we are dealing with (the sample of 900 pixels is at -8.28 EV, barely over black clipping).

Anyway, create more suitable raw images and we can talk about it. Until then it is not worth of discussion.

OP GaborSch Veteran Member • Posts: 7,203
There is no "black patch" in these images

If you can convince some of the frightened, unsecure Nikonians to create such images, then we can talk about that.

Iliah Borg Forum Pro • Posts: 25,841
Re: There is no "black patch" in these images

create such images

Black trap, and shooting OOF, scene occupying central 1/4 portion of the frame, 135mm lens or longer, f/8. But they will not avoid flare and resulting non-linearity.

-- hide signature --
Bernard Languillier Veteran Member • Posts: 4,672
Correct

Greg_WT wrote:

Gabor, did you take into account that Lloyd Chambers was writing
about 14-bit NEFs?

Testing with 12 bits files and extrapolating from these results the expected behaviour of 14 bits files belongs to the realm of modelization (define a model and use it to anticipate the output of non tested situations), not scientific experimentation.

The only scientific experimentation availalbe as of now are DxO results. They have a few very high key mathematicians in their scientific board (see this http://www.dxo.com/fr/corporate/home/committee ) and I do trust their findings.

Cheers,
Bernard

bobn2
bobn2 Forum Pro • Posts: 61,439
Re: Measurements

GaborSch wrote:

bobn2 wrote:

The difference has been observed and demostrated by Emil and others,
so at the moment it would seem that your measurement is the outlier

As I am measuring this instead of speculating about it, I have no
problem whatsoever.

I think Emil and the others were measuring rather than speculating , also. When one measurement disagrees with others, it's not speculation to say its an outlier. It's possible it's right and the others are wrong, but unless you can show the error in their methods, most of us are likely to take the majority view.

Read noise, sure. As for DR (more on that later) this depends on full
scale count (not necessarily the same as saturation) which can be
different from channel to channel. What this means is that a camera's
DR does indeed depend on the actual illuminant - hence Iliah Borg's
magenta filter trick

Sorry, you are mixing up two different issues.

1. Calculating the actual intensity, relative to the saturation
point: you need to know, when saturation occurs. I do know that
(though sometimes some camera copies contradict that). Anyway, don't
mix up the binary numerical range with teh saturation levels.

Good to know that.

2. The composition of the illuminant. I have a magenta filter, which
I use (extremely seldom) to increase the dynamic range of the
camera , NOT of the sensor .

Not exactly clear, actually. The top end of the DR of the sensor will depend on which of the colour channels tops out first. This in turn will depend on the illumination coming through the CFA, which is usually thought of as part of the sensor . Thus different sensors with different CFA characteristics will display their best DR with different illuminants. A basic problem with your method and all the others is that in the end you are using the sensor to measure itself. Kind of unavoidable without a lot of expensive lab kit (although I seem to remember John Sheehy came up with a proposal for a simple piece of kit which would do it), but you always need to be aware of the limitations caused by that.

Let's put it another way, how do you discover the 'saturation points'?

I am looking for raw files with saturated areas.

Good. Now I find this hard to square with your statement "It is not necessary to have any high exposure when measuring the DR; only the dark end is interesting."

I have not calculated the DR at all.

I find this difficult to square with your statement 'The result
proves, that the claim "two stops higher DR than the 5D2" is
ridiculous , like some other claims. In fact, the DR of the D3X is
max. 0.5 EV greater than that of the 5D2 at ISO 100, and at ISO 400
the 5D2 is already better.'

Back to my statement: the DR depends on the ISO, on the subject and
on the actual illumination, AND on your acceptance of noise.

I am not saying how high the DR of this or that camera is. I am
saying, that this or that camera creates a certain level of noise at
a certail level of illumination (measured on the raw channel)
. The
difference between the D3X and the 5D2 is FAR from the two stops,
even from one stop.

But that is a very different statement and says little about the DR. The bottom end of the DR is the read noise, and in your measurement that is swamped by the shot noise, which is about 2.5 stops higher than the read noise.

That's an SNR figure, not a DR. I've never been able to convince
myself that DxO's 18% grey noise value is a very useful metric

Not have I. I don't give a fig for how DxO measures the noise.

Good, we agree on that, then.

True, but it's always proportional to the ratio between full-scale
and read noise. All that happens is you might set your 'acceptable'
lower limit some amount above the noise floor

???

The bottom end of the DR ratio is set by what you think is the minimum acceptable signal distinguishable from the read noise. If you take it as at the read noise floor you get one figure. If you decide you want it one stop above the read noise floor, your DR measurement would be one stop less. So, if you had two cameras with one stop difference measured to the noise floor, the difference would still be one stop measured to one stop above the noise floor.

On many cameras 7EV from saturation is 6EV or more above the read
noise floor.
Shot noise will be about 3 stops greater than read noise at that level.
(Edit: I should have said 5 stops above read noise floor and shot
noise 2.5 stops above read noise)

Consequence?

Consequence is that differences in read noise will be masked by shot noise. As a numerical value, assume we have two cameras, one with a read noise of 6e-, the other with a read noise of 24 e-, if the FWC is the same, the former will have 2 stops better DR. On your measurement, you're measuring with a photon shot noise of say 136 e-. For the first camera, the observed noise will be sqrt(136^2 + 6^2) = 136, for the second the observed noise will be sqrt(136^2 + 24^2) = 138.

The effect is slight, but at the same exposure the shot noise in 900
pixels of 5DII will be lower than 900 pixels of D3x

Sorry, this is far out of the range we are dealing with (the sample
of 900 pixels is at -8.28 EV, barely over black clipping).

Yes, I said it was small.

Anyway, create more suitable raw images and we can talk about it.
Until then it is not worth of discussion.

Sorry, Gabor, I just don't think your measurements support the conclusions you make, they are quite consistent with the D3x have somewhere around two stops more DR than the 5DII. I think you are using a different definition of 'DR' to everyone else.

-- hide signature --

Bob

Fred Ferkel Contributing Member • Posts: 855
A 40D welding amateur

A 40D welding amateur should not make big boy type statements, it always turns out wrong.

bobn2
bobn2 Forum Pro • Posts: 61,439
Re: Measurements

GaborSch wrote:

The effect is slight, but at the same exposure the shot noise in 900
pixels of 5DII will be lower than 900 pixels of D3x

Sorry, this is far out of the range we are dealing with (the sample
of 900 pixels is at -8.28 EV, barely over black clipping).

Just to quantify this a bit, given a figure of about 66000 e- for the FWC of the 5DII (Clark, again) then -8.28 EV down from that gives 212 e-, for a shot noise of 15 e-. Assuming again the read noise figures of 6 e- for the D3x and 24 for the 5DII, we get an observed noise of 16 e- on the D3x and 28 e- for the 5DII. That's a ratio of 1.87 or 0.9 stop. Subtract the 0.1 stop that the area difference makes, and we're down to 0.8 stop, which is getting within range of your measurement, and is still consistent with the D3x having 2 stops more DR, if the FWC is the same or similar.

-- hide signature --

Bob

bobn2
bobn2 Forum Pro • Posts: 61,439
Re: A 40D welding amateur

Fred Ferkel wrote:

A 40D welding amateur should not make big boy type statements, it
always turns out wrong.
--

I wouldn't try to weld your 40D. It's made of magnesium, when the torch hits it it will go fzzzzzzzt.

-- hide signature --

Bob

Mel Veteran Member • Posts: 3,716
Re: The number one reason . .

This statement below sure makes it "seem" like he can be attributed to such. But you obviously know him so I will defer to your judgement that he is not.

"If you can convince some of the frightened, unsecure Nikonians to create such images, then we can talk about that."

If I had a D3X however, I would most certainly provide or at least ask him specifically what he needs to further this from his perspective.

I am still at a lost as to why he made his post to begin with if there is no intentional negative sentement toward Nikon? Is it then Dx0? But no matter, with his level of intelligence as far as knowing all of this, what end was/is served by posting such an unbased conclusion if he in fact needs this "created image" he speaks about to solidify his findings?

OP GaborSch Veteran Member • Posts: 7,203
Still measurements

bobn2 wrote:

I think Emil and the others were measuring rather than
speculating , also. When one measurement disagrees with others,
it's not speculation to say its an outlier. It's possible it's
right and the others are wrong, but unless you can show the error in
their methods, most of us are likely to take the majority view

Honestly, I don't give a fig for their "measurements", nor for the view of "most of you", for my measurements are good enough for me, beside being consistent. As soon as I see a contradicting measurement (which is not the same as a "statement"), I am ready to revisit my ones.

Not exactly clear, actually. The top end of the DR of the sensor
will depend on which of the colour channels tops out first

I am measuring raw channels .

Thus different sensors
with different CFA characteristics will display their best DR with
different illuminants.

A basic problem with measuring the DR on specific images under specific illumination is, that that result is totally useless in other setting or under other illumination.

A basic problem with your method and all the
others is that in the end you are using the sensor to measure itself

I am using the data gained from the sensor to measure the sensor's chartacteristics. I don't understand your objection. Should I use data from a specification sheet to support the specification sheet, or what?

Let's put it another way, how do you discover the 'saturation points'?

I am looking for raw files with saturated areas.

Good. Now I find this hard to square with your statement "It is not
necessary to have any high exposure when measuring the DR; only the
dark end is interesting."

I don't understand this problem (but really). I measure the saturation levels on any image created by the same model with the same ISO. Although it does happen, that some copies have different saturation levels, the deviation is always negligable from the perspective of dynamic range. (The deviations are important when determining if some patch really clipped ot not.)

Thus it is enough for me to find some other shots to measure the saturation levels; it is not relevant, that the actual images used in these demo are clipping or not.

Besides, all the above becomes irrelevant, when you regard that my demonstration used only the red channel, the saturation of which is very firm.

But that is a very different statement and says little about the DR.
The bottom end of the DR is the read noise, and in your measurement
that is swamped by the shot noise, which is about 2.5 stops higher
than the read noise

1. The DR is determined by noise . Not by "this noise" or "that noise" but simply be "noise".

2. You would have to show me hard proof before I accept that the shot noise is 2.5 times higher than the read noise in the range I used for the demo .

The bottom end of the DR ratio is set by what you think is the
minimum acceptable signal distinguishable from the read noise. If you
take it as at the read noise floor you get one figure. If you decide
you want it one stop above the read noise floor, your DR measurement
would be one stop less. So, if you had two cameras with one stop
difference measured to the noise floor, the difference would still be
one stop measured to one stop above the noise floor

I have not measured anything from the noise floor. I don't even have to know where the noise floor is. (In fact, I don't know that.)

Consequence is that differences in read noise will be masked by shot
noise

It is of no relevance whatsoever. Does the difference make one camera better than the other? (Beside the max. 1/3 EV difference I noted at the beginning?)

Sorry, Gabor, I just don't think your measurements support the
conclusions you make, they are quite consistent with the D3x have
somewhere around two stops more DR than the 5DII

I don't see this.

I think you are using a different definition of 'DR' to everyone else

I need to repeat it: I have not used the term DR in my demo . I used the terms "noise" and "intensity".

rhlpetrus Forum Pro • Posts: 25,860
Re: Clarification

Which files did you use (to Gabor)? Usually studio shots have much les DR than landscapes or than properly setups for measuring DR. Is their lighting controlled to present the DR avalilable in outdoors shooting?

I think posting such a claim w/o actually having control over the shots is a bit on the weak side, isn't it? I'd stay with DxO mark, at least for now, at least they did the whole testing setup, if mistaken, ok, but one should at least point where the falw could be.

-- hide signature --

Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/

Good shooting and good luck
(after Ed Murrow)

 rhlpetrus's gear list:rhlpetrus's gear list
Leica D-Lux (Typ 109) Nikon D7000 Nikon 1 V1 Nikon AF-S DX Nikkor 12-24mm f/4G ED-IF Nikon AF Nikkor 35mm f/2D +4 more
rhlpetrus Forum Pro • Posts: 25,860
full-well capacity Re: Demonstration

I didn't read the whole exchange, but I'm with you that one needs to know full-well capacity as well to actually compute DR, or am I completely off the mark here.

Can't understand how one measures DR by just looking at darker areas and computing noise there.

bobn2 wrote:

I'm not sure how you're inferring the DR from these SNR measurements,
perhaps you could give us the calculation? Usually, the sensor DR is
the ratio of saturation and read noise (or some multiple of it), i.e
the brightest feature the camera can resolve divided by the darkest.
Your measurements includes the photon shot noise in the sample, and
is also pixel referred (if your samples covered the same number of
pixels). Including the shot noise evens out the DR difference since a
big chunk of observed noise is in the photons and common to both.
Doing it per-pixel referenced tilts it a bit in the 5DII's favour,
because you're measuring a larger area of the image.

I suspect, if I've understood correctly, that the difference in your
measurement from others is because what you're calling 'DR' is
something different, namely the extra exposure that the 5DII needs
such that the resultant reduction in shot noise cancels out the
disadvantage in read noise, which unless I'm missing something isn't
the same thing (or maybe it is, need to think about this). In the
end, I can't see how you can calculate DR without knowing saturation.
--
Bob

-- hide signature --

Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/

Good shooting and good luck
(after Ed Murrow)

 rhlpetrus's gear list:rhlpetrus's gear list
Leica D-Lux (Typ 109) Nikon D7000 Nikon 1 V1 Nikon AF-S DX Nikkor 12-24mm f/4G ED-IF Nikon AF Nikkor 35mm f/2D +4 more
Iliah Borg Forum Pro • Posts: 25,841
Re: The number one reason . .

you obviously know him so I will defer to your judgement
that he is not.

I know Gabor's work on analysing raw images, and I know he is not biased. He may be mistaken sometimes, but that is for good.

If I had a D3X however, I would most certainly provide or at least
ask him specifically what he needs to further this from his
perspective.

I have, but I do not have 5D MkII; can't shoot them side to side; can't see how anything less than that is useful.

I am still at a lost as to why he made his post to begin with if
there is no intentional negative sentement toward Nikon?

Because 2 stops advantage of D3X over 5D MkII is a far stretch. It is also very dangerous, because it raises expectations and leads to frustration in the field/studio use. Add some jewellery and a black trap to the IR scene and you will know how it is in fact.

-- hide signature --
OP GaborSch Veteran Member • Posts: 7,203
THIS is the point

Iliah Borg wrote:

It is also very dangerous, because it raises expectations and leads to
frustration in the field/studio use

The vast majority of top photographers, who are the prospective buyers of the D3X, don't have any idea of the details we are dicussing here (the more expensive the camera, the less technical knowledge the owner has).

A much lower level example: the Canon 5D2. There were none on the shelves, long shipping time was the norm. Canonites were crazy about getting one a week sooner than possible, as it turned out partly with wrong expectation (but the relative law price offset that). After a few weeks of practical experience, the initial enthusiasm settled, not to say it sank. Now it is there, where it belongs.

I am absolutely sure, that the D3X is an exceptional camera, but even the best one can become cr@p for an owner if the expectation is too high.

Keyboard shortcuts:
FForum MMy threads