Nikon is NOT on crack - Initial D3x image quality is AMAZING!

[ejmartin]

This is the way the Canon 40D works, converted by ACR, NR set to zero. The subject is a grey strip wedge, very uniform.

The noise is lower if converted by DPP (NR set to zero). One could speculate about the reason: if DPP is doing some NR despite NR set to zero (ACR is doing so), or if ACR mistreats the 40D raw data. However, I am using ACR as are most photographers.

As far as I can tell DPP does not do any significant NR when it is turned off.  ACR is just a poor raw converter, especially at high ISO, when it comes to noise and resolution.

Of course I did. That is the suggested method for downsampling. Perhaps you choose a method, which is suitable to create a "proof", but I choose the best method for the purpose, namely downsampling with the least loss of details.

I like to hone a method that gives an accurately resampled image.  There are tools in common use, and there are better tools.  There is commonly used workflow, and there is better workflow.  

Oh well... At least I did not select the methods of creating the samples with the goal in mind to prove anything. I did this just like I am doing it normally.

Yes, and I suggested a major shortcoming of your workflow -- I pointed out that a judiciously chosen gaussian blur before resampling eliminates aliased noise in the downsampled image.  Guess you missed that.

Anyway, did someone claim that the noise increases by downsampling? Ray claimed, that the DR increases by downsampling. I find that nonsense, even though the noise can be decreased by downsampling.

You did, implicitly, by posting a file in which the noise at fixed spatial frequency went up after downsampling.   And no, the dynamic range at fixed spatial frequency is unchanged with proper downsampling.

[Ray]

Ray claimed, that the DR increases by downsampling. I find that nonsense, even though the noise can be decreased by downsampling.

No I didn't. If you have no detail in the shadows and no detail in the highlights, then clearly downsampling will not create detail where non previously existed.

I asked the following question.  Can we expect a higher DR and S/N than the D3 produces, when D3X files are downsized to 12mp?

If the D3X were to have the same DR and the same S/N noise as the D3, for example, at the pixel level (which I'd say is unlikely), then one could expect the downsampled D3X image to have at least marginally less noise than the D3, and by virtue of the lower noise resulting from the downsampling, one could also expect a correspondingly greater DR in relation to the D3 image, but not of course in relation to the original D3X image.

Is that now clear? This is how it seems to me, but I'm always happy to be proved wrong. I'm here to learn.  .

[grepmat]

Noise is not a number, it is a function of spatial frequency.

Noise can and is routinely defined as one single number. After all, how much spatial information is there in blackness?

There is sometimes some spatial components to fixed-pattern noise (e.g., "banding"), but that's quite a different thing.

Also, you speak as though you are an authority as to just what algorithms P.S., etc., uses. I suggest in fact you are just presuming.

Thanks.

[jani]

Noise can and is routinely defined as one single number. After all, how much spatial information is there in blackness?

There is sometimes some spatial components to fixed-pattern noise (e.g., "banding"), but that's quite a different thing.

Also, you speak as though you are an authority as to just what algorithms P.S., etc., uses. I suggest in fact you are just presuming.

Prior experience in the LL forum indicate that Emil generally has a pretty good idea about what he's talking about, and that he's good at assimilating information about how things work.

His list of publications seems to indicate that he may know a thing or two about statistics, mathematics, and possibly something about physics.

In other words, I think your mistrust is misplaced.

[grepmat]

Prior experience in the LL forum indicate that Emil generally has a pretty good idea about what he's talking about, and that he's good at assimilating information about how things work.

His list of publications seems to indicate that he may know a thing or two about statistics, mathematics, and possibly something about physics.

In other words, I think your mistrust is misplaced.

Thanks, but my qualifications are quite robust too, and I design digital camera sensors for a living. I'm not allowed to say for who, though.

Anyway, it's not a matter of trust, it's a matter of the credibility of some of his statements based on their content.

Cheers.

[Panopeeper]

If you have no detail in the shadows and no detail in the highlights, then clearly downsampling will not create detail where non previously existed

Removing the lens cap may help.

I asked the following question.  Can we expect a higher DR and S/N than the D3 produces, when D3X files are downsized to 12mp?

All right, you did not claim it but asked.

The issue is, tha the dynamic range in photographic sense is not the same as in engineering sense. It involves not only the lack of noise but the presence of details as well; otherwise the dynamic range could be increased by replacing all lower pixel values with a very low constant.

Of course, the level of exectable details is different between the D3 and the D3X due to the different resolution. Thus it is possible to increase the DR while descreasing the resolution (this is, what Emil encodes with "lower spatial frequency"). The problem is, that one could gain the impression in these threads that the image could be "enhanced" by downsizing, and that is the nonsense. The result will always be worse than the original.

by virtue of the lower noise resulting from the downsampling, one could also expect a correspondingly greater DR in relation to the D3 image, but not of course in relation to the original D3X image

Anyone buying a 24 Mpix camera for twice the price of a 12 Mpix camera with the intention of changing that back into another 12 Mpix camera is not fit for photography. This issue does not deserve any discussion.

[Ray]

Anyone buying a 24 Mpix camera for twice the price of a 12 Mpix camera with the intention of changing that back into another 12 Mpix camera is not fit for photography. This issue does not deserve any discussion.

Anyone who buys any camera with the intention of making prints only of a specific size in accordance with the native, uninterpolated resolution, is very inflexible.

[ejmartin]

Noise can and is routinely defined as one single number. After all, how much spatial information is there in blackness?

There is sometimes some spatial components to fixed-pattern noise (e.g., "banding"), but that's quite a different thing.

Also, you speak as though you are an authority as to just what algorithms P.S., etc., uses. I suggest in fact you are just presuming.

Thanks.

Noise has a spectral power distribution.  The standard deviation is an average over that power distribution, but there is much more information in the power distribution itself.  In particular, the power distribution dictates what happens to noise when a linear filter is applied to the image, as is done in resampling an image using any of the common filters -- Bicubic, Lanczos, etc.

The fact that there is no signal in blackness is an issue independent of the fact that there is a spatial frequency distribution of noise still present.  I was not referring to fixed pattern noise specifically, though that does indeed show up in the noise power distribution.

I did not and do not claim to be an authority about Photoshop's internal processing.  As far as I'm concerned it's a black box that one can perform experiments on.  One such experiment is to resample an image and examine the noise before and after (preferably the full power spectrum).  My comments were based on the analysis of such exercises, which indicate that PS Bicubic aliases noise into a downsampled image.  If you have contrary information, by all means present it.

[Panopeeper]

Anyone who buys any camera with the intention of making prints only of a specific size in accordance with the native, uninterpolated resolution, is very inflexible.

Ray, if I downsize an image for printing or for presentation on computer, then I am not doing that in order to make it like it were another camera or in order to compare it to another camera. In other words: I resize it according to the presentation medium, not according to another camera's specs.

[grepmat]

The issue is, tha the dynamic range in photographic sense is not the same as in engineering sense. Yada yada yada...

Dynamic range "in the photographic sense" is just an arbitrary and personal opinion and is hence useless during critical discussion.
Your main problem is that you want to force others to accept your very own special, undefined "definition" of what noise and dynamic range are. Sorry, that's not good enough.

[grepmat]

Noise has a spectral power distribution.  Yada yada yada

So? I can't summarize noise as X mv or counts, average RMS or sigma or full-width-half-maximum, etc? You think there's so much more information in a what is essentially a Gaussian distribution, for example, that one has to devote pages to it in a publication, or debate endlessly on your ill-defined special theories on a general photography discussion board? You are making me laugh!

[Panopeeper]

Dynamic range "in the photographic sense" is just an arbitrary and personal opinion and is hence useless during critical discussion

1. The "detail reproduction" part is not only arbitrary, but subjective as well; we have to live with that.

2. The "noise" part is not subjective, but arbitrary in the sense that the acceptable level of noise has to be declared, but it can not be declared universally.

Therefor, when I measure the dynamic range (only the noise aspect), I create a list of noise levels and the corresponding dynamic ranges per ISO setting; such lists from different cameras and different ISOs can be used for direct comparison.

Following is an extract from such a list, which is going to be incorporated in an Excel chart; this is not sorted yet. It has been measured on the raw channels from the shot of a Stouffer wedge, Canon 40D @ ISO 1600. The first number is the dynamic range, the second is the noise in percentage of the intensity.

6.61   25.2
5.12   13.0
5.24   13.2
6.77   26.5
5.30   13.4
5.41   14.7
6.94   31.2
5.49   15.1
5.59   15.7
7.12   32.0
5.64   15.3
5.75   17.6
7.16   31.7
5.69   16.5
5.80   17.3

[Ray]

Ray, if I downsize an image for printing or for presentation on computer, then I am not doing that in order to make it like it were another camera or in order to compare it to another camera. In other words: I resize it according to the presentation medium, not according to another camera's specs.

Nor is anyone else. That would be silly. I can't imagine why anyone would want to produce a print or presentation in order to create the impression the image was taken by a particular camera, unless one were playing a game, "Guess which camera was used for this shot?"

Cameras are tools. It's necessary to know what the strengths and weaknesses of different models and makes of cameras are, so one can choose sensibly in accordance with one's own photographic needs and interests.

It would be a bad decision to buy a particular model of camera for a perceived reason that it produced cleaner images at ISO 3200, if another camera with a higher pixel count could produce equally clean images when downsampled to the same size, but which in addition had the clear advantage of producing higher resolution images when downsampling was not required.

[grepmat]

By the way, I just realized that I have been conflating the prior two posters as one individual. My apologies for this.

[grepmat]

1. The "detail reproduction" part is not only arbitrary, but subjective as well; we have to live with that.

2. The "noise" part is not subjective, but arbitrary in the sense that the acceptable level of noise has to be declared, but it can not be declared universally.

No, you have to live with those arbitrary and subjective declarations. The rest of us will continue measuring and comparing camera performance based on normal scientific definitions and standards.

But here's a clue:

"Noise as a percentage of intensity" (not that what you mean is clear) should follow the statistics of shot noise - the fluctuations in the number of arriving photons. Wow, what a discovery! Oh, wait, didn't we get into a "discussion" about some bizarre interpretation of shot noise you were promulgating about a year ago? Right! Never-mind! Of course, this has essentially nothing to do with the performance of the camera, either.

[ejmartin]

So? I can't summarize noise as X mv or counts, average RMS or sigma or full-width-half-maximum, etc? You think there's so much more information in a what is essentially a Gaussian distribution, for example, that one has to devote pages to it in a publication, or debate endlessly on your ill-defined special theories on a general photography discussion board? You are making me laugh!

Setting aside pattern noises (fixed or otherwise), the noise coming off the sensor is highly uncorrelated and indeed close to Gaussian.  However, the RAW conversion process and in particular demosaicing (and noise reduction inherent to some converters) make it something quite different, if one looks at the spectral power distribution.  A measurement of the standard deviation alone will not dictate what happens to the noise, DR, S/N etc of an image under resampling.

The following three patches have the same noise as measured by the standard deviation:



The "grain size" of the noise is related to the shape of the noise spectrum



and dictates what happens to the std dev under resampling -- "coarse" grain (red) is relatively unaffected by resampling, since it has little noise power near Nyquist.


I'm not sure why you feel the need for condescension.

[grepmat]

However, the RAW conversion process and in particular demosaicing (and noise reduction inherent to some converters) make it something quite different, ...

The RAW conversion process has little to do with the performance of the sensor or the camera (if RAW is truly RAW). We all know how drastically different the results can be between the various RAW converters.

If you instead want to comment on the relative performance of different RAW converters, then have at it. Indeed, I feel that this is worthwhile, as simple noise results post-conversion (e.g., DPReview plots of JPG noise) can be almost meaningless if one wants to evaluate the performance of the underlying hardware.

But if you want people to accept your methodologies, I suggest defining them mathematically and keeping it simple. If you want to discuss frequency spectra, etc., remember how difficult it is for people to properly read MTF tables, for example. Ideally, you would deliver one figure of merit that represents the over-all performance of that camera+RAW converter.

I'm sorry if I was condescending. As I mentioned, I was confusing you and panopeeper.

[Panopeeper]

No, you have to live with those arbitrary and subjective declarations. The rest of us will continue measuring and comparing camera performance based on normal scientific definitions and standards

I'm afraid the rest of you will be a quite small group, as the vast majority of photographers prefers to decide subjectively, if the detail reproduction is good enough for her/him, based on the individual circumstances.

"Noise as a percentage of intensity" (not that what you mean is clear) should follow the statistics of shot noise - the fluctuations in the number of arriving photons. Wow, what a discovery! Oh, wait, didn't we get into a "discussion" about some bizarre interpretation of shot noise you were promulgating about a year ago? Right! Never-mind! Of course, this has essentially nothing to do with the performance of the camera, either.

I think you lost the orientation in the thread.

[grepmat]

I'm afraid the rest of you will be a quite small group, as the vast majority of photographers prefers to decide subjectively, if the detail reproduction is good enough for her/him, based on the individual circumstances.

I have no problem with people making subjective judgments. I do the same every day. I only scoff at people that insist it is necessary to co-mingle arbitrary and subjective factors in their "scientific" analysis.

[ejmartin]

The RAW conversion process has little to do with the performance of the sensor or the camera (if RAW is truly RAW). We all know how drastically different the results can be between the various RAW converters.

If you instead want to comment on the relative performance of different RAW converters, then have at it. Indeed, I feel that this is worthwhile, as noise results post-conversion can be almost meaningless if one wants to evaluate the performance of the underlying hardware.

I agree with the sentiment that noise performance is best discussed at the level of the RAW data.  However, even then I do think it important to keep in mind that noise is frequency dependent, even though it is uncorrelated.

But if you want people to accept your methodologies, I suggest defining them mathematically and keeping it simple. If you want to discuss frequency spectra, etc., remember how difficult it is for people to properly read MTF tables, for example. Ideally, you would deliver one figure of merit that represents the over-all performance of that camera+RAW converter.

For (uncorrelated) noise in the RAW data, I think a good figure of merit is the slope of the noise power spectrum; equivalently the noise at a fixed reference frequency; equivalently the standard deviation at the pixel level divided by the number of vertical pixels in the frame.  When the noise is spatially uncorrelated these are all the same thing up to fixed known dimensionless constants.  The advantage of such a figure of merit is that it is scale invariant; resample the image and the figure of merit remains unchanged, as indeed it should be for fixed image output dimensions.  Noise should be referred to frame size, as one does for resolution, not given as an absolute number at the pixel level.  I think we both agree that the scalings involved at the RAW level are trivially performed -- just rescale the pixel level noise by the appropriate power of the pixel count.

Post conversion, however, I don't think one can boil things down to one number.  Perhaps a few would do.  But now the above three measures of noise are no longer the same -- the slope of the noise power is the same as the noise at a reference scale, provided both are in the linear regime of the power spectrum; but the std dev is very much affected by what the converter did to near-Nyquist data.  I think I would prefer one of the first two over the third, since they still refer to fixed output dimensions are relatively immune to what the converter is doing, and add as another figure of merit the "grain size" imparted by the converter perhaps as measured by the point where the spectrum starts to deviate from linearity.