Nikon D4 Read Noise Measured and Compared

Marianne Oelund said:

It's hard to imagine that they would have intentionally sacrificed some QE in order to provide a base ISO of 100 at a more modest FWC, but who really knows?

Click to expand...

I think that's exactly what happened. Which makes it logical that again a D4s will pop up in about two years time.

The one thing that's puzzling me is the D800 name. It's pretty much sure that it replaces the D3x, since it is expected to come with the 36 Mp sensor.

If they stick to a lo price FX camera ( like the current D700), theres a name collision. Either they drop that strategy and make FX an exclusive product, or the keep the low cost FX and name it D700s.

The first option means there will be an upgrade in the Dx range for sure, one with lo noise at higher isos (D400 16 Mp from sony) and one with hi Mp (D400x 24 Mp from sony) and pro video as a contrast (similar to the FX strategy).

The second option will lead to a stripped down D4, renamed as D700s, with a body similar to the current D700, less fps, lower iso range (which will not make much of difference in every day use) and less professional video options. A D400 DX loaded with the 24 Mp from sony and pro video options will be the top end of aps-c sensors.

In any case, there's a lot to come.

lock

David314 said:

Iliah Borg said:

For raw data; white balance pre-conditioning is applied to raw data.

It is a Nikon thing since D2x.

Click to expand...

i've never understood what that really means as the raw data itself is not a function of the white balance setting on the camera

Click to expand...

Correct. These are simply fixed scalings which are always applied to the red channel and blue channel ADC outputs, and they do not depend on any camera settings like WB. They differ slightly for each sensor type, so it appears to me that Nikon only do this to "equalize" sensor channel responses between camera models, which then would allow them to use common firmware modules.

The scale factors are just a bit more than unity, e.g. about +12% and +16% in the case of the D7000, if memory serves correctly.

David314 said:

there is certainly non doubt that raw data is not truly raw, one need only shoot at f1.4 to see that - at least outside the optical black sensors

Click to expand...

I'm not sure exactly what you're referring to. Any specific examples?

Marianne,

Marianne Oelund said:

These are simply fixed scalings which are always applied to the red channel and blue channel ADC outputs, and they do not depend on any camera settings like WB.

Click to expand...

FWIW.

I often see slight scaling in the green channels as well.

I occasionally see different scaling with the same exact camera so I'm not sure they are really fixed but they do vary very little.

I think I have seen some variation between cameras of the same model as well.

Regards
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

bclaff said:

bclaff said:

the correlation between masked-pixel noise and image-pixel noise needs to be established anew.

Click to expand...

I don't think there's any real reason to expect that the "optical black" approach is no longer valid; but I agree it's a worthwhile future exercise.

Regards,
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Click to expand...

Yes hi Bill, Agreed with you and Marianne on future measurements, wish I had a D4 to play with .

All I'm saying is, if a new method has the attractive advantage of ease of use, etc. then my antenna goes up immediatly to ask questions along the lines of accuracy and correspondence to the traditional methods. If the method is solid, then it should be able to stand up to my challenges! I'm still trying to wrap my head around your D7000 optical black data:

http://www.luminous-landscape.com/forum/index.php?topic=60403.msg486844#msg486844

I mean, which set of columns gives the "real" read noise? Have people given you theories about this?

Thanks,

Chris

Marianne Oelund said:

Marianne Oelund said:

there is certainly non doubt that raw data is not truly raw, one need only shoot at f1.4 to see that - at least outside the optical black sensors

Click to expand...

I'm not sure exactly what you're referring to. Any specific examples?

Click to expand...

i read a nice article about this about a year ago but couldn't find it again

so example

raw histograms with two shots of the same scene f1.4 and F4

Chris,

cpw said:

I'm still trying to wrap my head around your D7000 optical black data:

http://www.luminous-landscape.com/forum/index.php?topic=60403.msg486844#msg486844

I mean, which set of columns gives the "real" read noise?

Click to expand...

Because of they way they match numbers gathered in the traditional way, the columns away from the effective area are the ones with the "real" read noise data.

(The columns with the "unknown" values are immediately adjacent to the effective area.)

cpw said:

Have people given you theories about this?

Click to expand...

I have no theory at this time that's why I was reaching out to others for ideas.
Perhaps I should make a similar inquiry here at dpreview.

Regards
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Marianne Oelund said:

But that's only if we have enough image data to evaluate FWC, which allows us to determine the "gain" figure (gain = FWC/2^14 for 14-bit ADC). Lacking that, we can only work in ADU for everything, simply taking the viewpoint of the RAW file data values alone. Bill has done this to come up with dynamic range figures from the read noise in ADU.

Click to expand...

This is probably a silly question but if DR is defined as the range of LVs a sensor can capture between the noise floor (1:1 SNR maybe) and saturation, how can DR be calculated without gain and FWC? A DN near 2^14 represents the ADC-converted value of light saturation and sensors are ^2 linear, but if DN were enough to define saturation then every 14-bit sensor would have a FWC representing 2^14 photographic stops (LVs) and so DR would be solely determined by the noise floor?

For example let's take two sensors, both with a base ISO read noise of 4e- but sensor 'A' has a FWC of 128k e- and sensor 'B' has a FWC of 64k e-. Using 1:1 SNR as the noise floor, doesn't 'A' have a DR of 16 LVs and 'B' have 14? 'A' will have fewer tonal gradations per LV vs 'B' over the 2^14 ADC range since it's compressing 2^17 down to 2^14.

I know Bill's chart is labeled "engineering DR" but I always thought that simply meant a DR with an objective and standardized noise floor (like 1:1) as contrasted with a more subjected definition where detail and color rendition would be factors establishing a "subjective DR".

Marianne Oelund said:

Marianne Oelund said:

For image sensor technologists:

Read noise is measured in electrons r.m.s. and comparison among sensors/cameras in bits (aka ADU) is only useful when the conversion gain (uV/e-) and ADC conversion (Volt/bit) is known.

Click to expand...

We're doing the same thing, but combining the two separate measurements into one, and we tend to think in the inverse relationships (because we're a quirky bunch): Instead of knowing the separate figures for e-/uV and uV/ADU, we just work with the product, e-/ADU, which we simply call the "gain" (but which, properly speaking, is actually inverse gain). Then it scales inversely with the ISO setting. Typically, we express all noise results in both e- (which we denote as "input-referred") and ADU (which we denote as "output-referred"), because they each lend useful perspectives.

Click to expand...

OK, but I did not see e- rms in the linked data. Did I miss it? Comparing among different sensors/cameras based on ADU rms may not yield sensible comparisons. I get the sense you know this but perhaps not others.

Marianne Oelund said:

But that's only if we have enough image data to evaluate FWC, which allows us to determine the "gain" figure (gain = FWC/2^14 for 14-bit ADC).

Click to expand...

I guess you are assuming someone at Nikon set the top of the ADC to the FW value. I think it is set higher for margin sake.

Lacking that, we can only work in ADU for everything, simply taking the viewpoint of the RAW file data values alone. Bill has done this to come up with dynamic range figures from the read noise in ADU.

Marianne Oelund said:

Marianne Oelund said:

Dynamic range is the 20 log (Smax/Nd) where Smax is the maximum signal that can be captured (often linearly captured) and Nd is the noise level in in the dark. Smax and Nd need to be in the same units, ADUs, electrons, volts, etc.

Click to expand...

Yes, here we are using strictly ADU, although we prefer to think in terms of powers of 2 (stops) instead of dB.

Click to expand...

OK, that makes sense. So a DR of 10 means 10 stops for you, or 60 dB, right?

bclaff said:

But ADUs (a/k/a/ DNs) are more useful in some ways because this is the "raw" data we actually get as camera users.

Click to expand...

OK, but obviously this is for a particular camera and you cannot meaningfully compare different cameras.

bclaff said:

bclaff said:

Dynamic range is the 20 log (Smax/Nd) where Smax is the maximum signal that can be captured (often linearly captured) and Nd is the noise level in in the dark. Smax and Nd need to be in the same units, ADUs, electrons, volts, etc.

Click to expand...

Yes.
I use SNR for a shorthand that is literally S/N.

Click to expand...

Yes, but as long as you are using the noise in the dark under the same settings then it is ok.

Optical black pixels could potentially have a difference conversion gain (uV/e-) then the optically active pixels, btw. It depends on how they are made black.

bclaff said:

I find db a confusing unit since for everything exposure related we use log2 so I use that instead.

Click to expand...

ok, sure.

bclaff said:

Could you comment on the shape of the D4 curve as opposed to it's predecessors?
As well as how the curve progress upwards over time?
(The links are in my initial post)

Click to expand...

The shape looks reasonable. The noise has two sources, something probably after the gain, and some before the gain. The former is not affected by gain and dominates the noise at low gain. At higher gain, the earlier part of the signal chain noise is amplified and comes to dominate the total noise. That is just a general interpretation and perhaps there are others.

When you say over time I guess you mean from generation to generation. Seems reasonable that they would try to increase FW and reduce read noise in succeeding generations. Since FW is independent of gain (ISO setting) the curve looks like an upside down version of the read noise curve. It is a good but not mindblowing improvement, but I don't know the difference in pixel sizes etc. Looks a little better than a factor of 2 between FW increase and read noise decrease. There is nothing here that does not mirror the technological improvements in camera phones sensors. Sorry I am blase about a one stop improvement in DR. I am happy tho that it apparently makes a big difference from the end user perspective!

Eric,

Eric Fossum said:

OK, but obviously this is for a particular camera and you cannot meaningfully compare different cameras.

Click to expand...

I think normalizing by photosite area is valid to compare cameras.

But I prefer "real world" tests like my Photographic Dynamic Range (like DxoMark) to compare cameras.

Eric Fossum said:

Optical black pixels could potentially have a difference conversion gain (uV/e-) then the optically active pixels, btw. It depends on how they are made black.

Click to expand...

AFAIK optical black photosites are "made black" solely by blocking exposure to light.
(But I am sometimes wrong in my assumptions!)

Regards
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Adam,

DR is simply a ratio usually expressed as db but in this type of analysis often expressed as log2.
As such, the units (ADU versus e-) are irrelevant.

For me "Engineering DR" is the traditional DR that uses the noise floor as the denominator.

I distinguish that from "Photographic DR" where I choose I higher denominator based on a S/N criteria.

Regards
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Lets stop all this speculation and wait for the camera to be available before any comparisons/reviews/etc. comparing x to y, nikon to canon, d4 to 1dx, d4 to d3s, the d800, d5.....

A distinguished visitor has just joined the thread, so there is something to be interested in. Several of us are also deciding whether or not to pre-order this camera. Since all we have to go on are the sample images so far in order to make this decision, we're grateful for the assistance.

Luke Kaven said:

A distinguished visitor has just joined the thread, so there is something to be interested in. Several of us are also deciding whether or not to pre-order this camera. Since all we have to go on are the sample images so far in order to make this decision, we're grateful for the assistance.

Click to expand...

No serious person would ever base a pre order decision of a new not yet released camera on amateurish samples posted on the net.

Thank you. Finally someone with reasoning. I don't have to be a famous photographer o have reasoning Luke.

Actually, if you can read, you would notice that some valubale information can be ascertained. This is a gear forum, discussing a new camera, with some very prestigous theorists mixing in. If you don't like it butt out. I for one find it fascinating and educational.

bclaff said:

Adam,

DR is simply a ratio usually expressed as db but in this type of analysis often expressed as log2.
As such, the units (ADU versus e-) are irrelevant.

For me "Engineering DR" is the traditional DR that uses the noise floor as the denominator.

I distinguish that from "Photographic DR" where I choose I higher denominator based on a S/N criteria.

Click to expand...

Thanks Bill. What then is the numerator? Is engineering DR just a measurement of internal noise without respect to input signal?

Adam,

Horshack said:

What then is the numerator?

Click to expand...

It ought to be the saturation or white point.
For my analysis I used the ADU maximum.
For 14-bit I use 16384 (slightly careless since 16383 is really the maximum).

The real white point might be somewhere around 16000; this is about a 1/30th log2 difference, not too important.

Horshack said:

Is engineering DR just a measurement of internal noise without respect to input signal?

Click to expand...

As you said eariler, it's where signal equal read noise; below which you could say there is no measurable signal since the signal would be than the noise.

--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

bclaff said:

bclaff said:

What then is the numerator?

Click to expand...

It ought to be the saturation or white point.
For my analysis I used the ADU maximum.
For 14-bit I use 16384 (slightly careless since 16383 is really the maximum).

The real white point might be somewhere around 16000; this is about a 1/30th log2 difference, not too important.

Click to expand...

But a quantized DN of the ADC (near 2^14) can refer to any number of actual captured photons, correct? A 2^14 DN on the D3s refers to near 90k e-, whereas 2^14 DN on D3/D700 is near 58k e-. So isn't the read noise and resulting engineering DR derived from it just an internal noise measurement without respect to input signal?