Michael Reichman's comment on D3 @ High ISO

[Alain Briot]

Here's a similar comparison of the D3 at ISO 6400. Simple, literal RAW conversion on the left; embedded Nikon JPEG on the right:

John,

Thank you. Does the Raw have noise removal applied automatically by the camera?

Also, what is the purpose of the jpeg? I shoot only in raw and never use jpeg so the appearance of the jpeg is of no consequence to me.

 

[John Sheehy]

John,

Thank you. Does the Raw have noise removal applied automatically by the camera?

Also, what is the purpose of the jpeg? I shoot only in raw and never use jpeg so the appearance of the jpeg is of no consequence to me.

I don't think the camera performs any noise reduction on the RAW. I binned dark, flat areas to 2x2, 4x4, and 8x8 and the noise reduced just as would be expected with real, natural noise. When noise is filtered, it does not reduce with the smaller binning (2x2) like it should.

The original D2X comparison was not for you per se; it was a demonstration of the real chromatic noise underlying JPEGs and conversions that seem to be lacking in chromatic noise. My RAW version is not what you're going to get with a converter; converters by default will darken the shadow areas, as they are bright from noise in the RAW. Converters will generally filter out most of the chroma noise.

 

[Alain Briot]

I meant that you seem to be impressed by NR, in leiu of the natural look.
If you like the NR and it looks pretty much how you would do it in PP, then that's nice, but one should never mistake the lack of obvious noise in an image as the result of low-noise technology. Nikons have always had a more aggressive removal of chromatic noise at all tonal levels, and total desaturation and blurring in the deepest shadows, by converters. The RAWs have always been chock full of lots of chromatic noise.

John,

OK, so I'm still not clear about your previous remarks regarding NR (see above).

Were you referring to jpegs and did you mean that the D3 performs noise reduction when it saves an image to jpeg format? A number of people mentioned that "there was no way to turn off NR reduction in the D3/D300." Were they referring to jpeg conversion?

Or, are you making the point that all Raw Converters remove chromatic noise and that we are therefore not necessarily aware of the actual chroma noise in a Raw file ?

Also, what is binning?

 

[John Sheehy]

John,

OK, so I'm still not clear about your previous remarks regarding NR (see above).

Were you referring to jpegs and did you mean that the D3 performs noise reduction when it saves an image to jpeg format? A number of people mentioned that "there was no way to turn off NR reduction in the D3/D300." Were they referring to jpeg conversion?

I don't know what other people are referring to, but all in-camera JPEGs have NR to some degree or other. Nikon has always had a very aggressive policy against chromatic noise, to the point of removing saturation completely in the deepest shadows (even of subjects). If you copy and paste the images into PS, you can use the Info tool to inspect the dark areas, and you will see in the JPEGs that many areas like the table-top have 0 saturation throughout. Commercial converters and manufacturer's converters lean this way too. Would you want people to use your RAW converter, and exclaim that the RAW converter does a worse job than the camera jpegs?

Or, are you making the point that all Raw Converters remove chromatic noise and that we are therefore not necessarily aware of the actual chroma noise in a Raw file ?

Also, what is binning?

Yes, it's just a matter of degree.

Binning is when you add pixels together to make super-pixels, but less image resolution. Perfectly random noise should double, as a standard deviation, when you add 4 pixels into one (but the noise actually halves, relative to the new 4x signal). When I did that with crops of smooth areas of one color channel of the D3 file, the noise changed as predicted, which means that it was not filtered to any significant extent.

 

[Martin Kuivenhoven]

I'm using a Pentax K10D for now. I'm waiting on the announcements from Pentax coming up next week to determine my upgrade path. I don't think a new Pentax camera is going to do it for me though. I'm pretty disappointed with their flash system and from what I understand, there's none better than the Nikon system, so...

However, I'm trying to practice delayed gratification at this point!

Seems Benjamin Kanarek is playing with the new Pentax toy as we speak .. maybe he can offer some insight soon

 

[Edward Bussa]

While I'm waiting for the announcements on the 25th, are there any threads you could point to?

It seems Benjamin has stopped posting here to a large degree, but I am aware he has continued posting elsewhere... Does he allude to a Pentax 645 or the K20D ?

 

[Alain Briot]

Binning is when you add pixels together to make super-pixels, but less image resolution. Perfectly random noise should double, as a standard deviation, when you add 4 pixels into one (but the noise actually halves, relative to the new 4x signal). When I did that with crops of smooth areas of one color channel of the D3 file, the noise changed as predicted, which means that it was not filtered to any significant extent.

Thank you.

So, is binning synonymous with downsampling? How do you select the number of pixels you "bin"? In Photoshop you can only select the type of downsampling, not how many pixels are used. Do you use a different software?

Regarding NR I only shoot raw, never jpegs, so I am not concerned with camera-applied NR. However, it is interesting to know that NR is done on in-camera jpegs, something that I ignored.

 

[John Sheehy]

Thank you.

So, is binning synonymous with downsampling?

Not exactly. Downsampling should use filtering on the image before it is applied, and uses some original pixels more than once to calculate different output pixels; it is a softer effect, while binning strictly combines distinct multiple pixels, and they are only used once. Binning can lead directly to aliasing.

How do you select the number of pixels you "bin"? In Photoshop you can only select the type of downsampling, not how many pixels are used. Do you use a different software?

Iris, the program I use to look at RAW data has a literal binning function. PS can emulate binning by using the "Pixellate|Mosaic" function, but the difference is that the resulting image will not get smaller, and the depth of precision won't increase as it does when you just add the pixels together. PS divides the results by the number of pixels binned together. Truly binned pixels get brighter because of the values adding together. The Mosaic filter's slider determines the number of pixels on a side of a square to be binned; IOW, if you select "5", 25 pixels in a square will be added together and then divided by 25, and the result will be a 5x5 pixel square of solid color.

 

[Andrew Rodney]

A 12.8MP would only give better results than a 12.8MP crop from the 21.1MP. You don't get less image noise by having less and bigger pixels in the same size sensor. This is a very popular but totally unfounded myth.

I think the only fair way to compare would be to sample down the larger image to that of the smaller image. That alone will reduce noise as you're talking several existing pixels (represented as noise in shadows) and sampling them down to less, the result is less noise.

 

[Martin Kuivenhoven]

While I'm waiting for the announcements on the 25th, are there any threads you could point to?

It seems Benjamin has stopped posting here to a large degree, but I am aware he has continued posting elsewhere... Does he allude to a Pentax 645 or the K20D ?

No specific reference to a product/type as far as I am aware: http://www.pentaxforums.com/forums/...75-getting-new-toy-onmonday-5.html#post157899

 

[John Sheehy]

I think the only fair way to compare would be to sample down the larger image to that of the smaller image. That alone will reduce noise as you're talking several existing pixels (represented as noise in shadows) and sampling them down to less, the result is less noise.

That would do well to compare a 12MP from each, but it is rather unfair to the camera with more pixels to give up its resolution in the comparison. Another way is to upsample both, and step back from the monitor, or print both large.

Unfortunately, monitors (and video cards) give an unfair advantage to low-MP images with their coarse resolutions. If monitors has such fine dot pitches that we upsampled images to view them fullscreen, we would appreciate high-MP images much better, without any need to downsample them to hide noise.

 

[Bart_van_der_Wolf]

That would do well to compare a 12MP from each, but it is rather unfair to the camera with more pixels to give up its resolution in the comparison.

I agree. Why disadvantage one camera more than another? It would never be considered a fair comparison, untill one only focuses on noise at the pixel level, which is not all that relevant for output. The problem with such 'comparisons' is the apples versus oranges aspect.

Another way is to upsample both, and step back from the monitor, or print both large.

I've heard that argument before, but it is only valid if, and only if, the specific interpolation algorithm doesn't favor small increments versus larger increments in size (or vice versa) for all spatial frequencies (thus also for noise which has a (specific) spatial frequency spectrum), which is not a given for all algorithms. I think only a more elaborate test which includes human visual perception would suffice. Unfortunately such a test will depend on the specific viewing conditions. Therefore, it seems more useful, to me anyways, to focus on the particular aspect (resolution or noise) at the native resolutions. If and how it scales with interpolation for final viewing conditions, then becomes a separate resampling and output method/size issue.

To put it in simple terms, resolution doesn't improve with interpolation and noise scales differently depending on its spatial frequency for the specific output and viewing size. If resolution is more important for the purpose, the choice is simple. If noise is (more) important for the purpose, there may be ways of improving noise in postprocessing that are sufficient for the purpose.

Bart

 

[John Sheehy]

I've heard that argument before, but it is only valid if the specific interpolation algorithm doesn't favor small increments versus larger increments in size (or vice versa) for all spatial frequencies (thus also for noise which has a (specific) spatial frequency spectrum), which is not a given for all algorithms.

I really don't see how using bicubic, for instance, at 220% on one image and 280% on another could possibly skew any comparisons to any degree that would outweigh the benefits of viewing at the same size, and without loss of resolution.

 

[Alain Briot]

Iris, the program I use to look at RAW data has a literal binning function.

John,

Very interesting. Do you have a link where I can see what Iris is all about? I did a search and found a free astronomical software package called Iris: http://www.astrosurf.com/buil/us/iris/iris.htm

Is this the one?

 

[Bart_van_der_Wolf]

I really don't see how using bicubic, for instance, at 220% on one image and 280% on another could possibly skew any comparisons to any degree that would outweigh the benefits of viewing at the same size, and without loss of resolution.

Okay, so you challenge how (e.g. ?) Bicubic (who uses that, inspite of better algorithms, anyways?) interpolation at 220% versus 280% outweighs viewing benefits, right? Let's define Bicubic resampling, Photoshop's version of it that is, or the mathemathical version, or ACR's, or ...?, and at which output magnification? There are too many variables, even if we disregard the specific noise spectra from the different cameras, to arrive at a uniformly informed opinion. I would agree with you if you challenged that.

Let's separate all variables, and put weights to them (for different output purposes), to make an objective (although informed) choice.

Bart

 

[Bart_van_der_Wolf]

Is this the one?

Yep, IRIS is reguarly used as a utility, because astronomical imagery is very demanding in getting each and every photon to contribute to the final result (photons are a scarce resource there, and we can learn a lot of fundamental stuff from that different photographic discipline). However, the binning (of multiple sensel 'scores') has several variations that could result in slightly different results (from an astrophotographic perspective), depending on the exposure time and capture device. The most common versions are averaging, or adding, sensor element 'scores'.

Bart

 

[John Sheehy]

John,

Very interesting. Do you have a link where I can see what Iris is all about? I did a search and found a free astronomical software package called Iris: http://www.astrosurf.com/buil/us/iris/iris.htm

Is this the one?

It's the one, but I wasn't talking about binning to suggest it. I mentioned binning, only because it is one way to tell if noise has been filtered by software. Binning, in and of itself, for the purpose of reducing noise, is false ecoomy, IMO, unless your optics are extremely oversampled, in which case binning won't lose much resolution.

You have to know a bit about RAW files to make much use of them in IRIS (and about the program, too, and it is a bit quirky). It is not a useful RAW converter in the general sense. It can display RAW data in a flexible manner, but never makes a true full conversion, like the commercial converters do, but you can try to get close with manually applied saturation, contrast/gamma curves, WB, USM, etc. I do like to view and compare RAWs in it, however, as it treats the RAW files from different cameras the same way, so if, for example, you took the same shot with two different cameras (everything else equal), and viewed them in IRIS with the levels tweaked to make them both the same brightness, you can see which has better absolute signal to noise ratios. This is the big shoot-out, which, it seems, is never going to happen, because no one who shoots camera comparisons ever provides the RAW and uses the same absolute exposure.

 

[John Sheehy]

Okay, so you challenge how (e.g. ?) Bicubic (who uses that, inspite of better algorithms, anyways?) interpolation at 220% versus 280% outweighs viewing benefits, right? Let's define Bicubic resampling, Photoshop's version of it that is, or the mathemathical version, or ACR's, or ...?, and at which output magnification? There are too many variables, even if we disregard the specific noise spectra from the different cameras, to arrive at a uniformly informed opinion. I would agree with you if you challenged that.

Let's separate all variables, and put weights to them (for different output purposes), to make an objective (although informed) choice.

I'm basically saying that for the stated purpose for which I am using the upsampling, I really don't care at all about minor artifacts, because they are relatively insignificant. I don't pixel-peep these comparisons; I squint or step back far from the monitor. I am not going out of photoshop to use a better algorithm, just to look at images on the screen at about ~250% and compare them. I am only interested in better algorithms when I really have to stretch things big, permanently, and the artifacts would become apparent.

 

[Alain Briot]

It's the one, but I wasn't talking about binning to suggest it. I mentioned binning, only because it is one way to tell if noise has been filtered by software.

John,

Thank you. Definitly quite different from Photoshop, Lightroom, DXO, and the other software I use.

ALain