I think you stretched the concept over the limit. We can practicall turn off white balancing, we can turn off contrast and saturation, but we can't turn off the de-mosaicing.

The LED light is practically single wavelength; as such, it passes one (or two) filters to a high degree, while the other (or others) let only a relative small portion through, although it appears measurably in all three colors, see spectral response. Therefor you can clip a channel with a fraction of the values in the other channels. However, the de-mosaicing will "distribute" the pixel values between the neighbours.

You mentioned "highly saturated". Led and laser lights are highly saturated, for sure - but highly saturated what? There are many different color within for example the range of "red" in the spectrum, but there is only one saturated red in the color space.

Consequently the highly saturated color you used may appear as a composite color from the point of de-mosaicing.

I agree that this would be far preferable to what we have now. But for colors that fall within sRGB, UniWB delivers a closer correlation between camera histogram and RAW than any other alternative.

I agree this is an extreme case; that's why I tried it--to see how well UniWB worked in extreme circumstances. I don't think that this invalidates the usefulness of UniWB in most circumstances, though.

I don't see how demosaicing would be relevant. You would get the same relatively saturated capture if R, G, and B were co-located.

I stated, that DoF depends on the lens, but not on the cropping; this is a fact. It is not useful to mix up concepts.

One can bring the sensel size in the equation as well; that again has nothing to do with cropping. I had similar discussion with paper tigers, whose measure of image quality is how large an image can be printed. They will debate endlessly over side-issues, like sensel size, number of pixels, etc.

One needs to see the issues more abstract, otherwise we can not conduct a factual discussion. Another such fruitless discussion was, that some people stated that the perspective depends on focal length and cropping. Plain BS. You can state, that with a different focal length you have to go to a different distance. Right, that changes the perspective, not the focal length.

Re panos in architectural photographing: it is unquestionable, that you can achieve a very good result with a wide angle lens or tilt and shift, but do you have such wide lens?

http://www.panopeeper.com/panorama/GameRoom1.jpg



There are many stitchers for casual panomakers. For those, who are serious about it, there is only one: PT (and its descendants, working on the very same principle). However, PT is not for human consumption, you have to have a good user interface. I prefer Panorama Tools Assembler, others prefer PTGui, and there is Hugin, but not on all platforms.

You certainly would not get the same, because the colors of a single tri-color pixel would have to undergo "only" a color space transformation, while the de-mosaicing has to work with several neighbouring pixels, where the distance between them is relevant too (and the result undergoes the color space conversion).

Nevertheless, there is no saturated color on raw level. You can not find any wavelength, which would be totally filtered by two channels. Theoretically, one wavelength per filter would be acceptable, but this irrelevant, and anyway there are no such filters.

So, you can have many different color compositions from different saturated colors (single wavelengths), which have to be transformed in different RGBs; only one of them may be transformed in a fully saturated color.

You guys are getting way off base here when you talk about red lasers, LEDs, and the highly saturated or monochromatic colors produced by these devices. It is not important to capture such colors since they do not occur in nature (or only very rarely) and can not be reproduced on any current display or printed. What is important for photography are the real world surface colors, which are discussed here.

Have you ever shot at a dance club, or concert? Or a car show? Highly saturated colors outside Adobe RGB aren't that uncommon. Not natural, perhaps, but that doesn't mean you'll never find them in a photograph. Knowing what to do when you do is of more than merely academic interest.

This has been discussed many times. If you want to capture all real world surface colors, you should render into ProPhotoRBG. Also, saturation clipping is easily seen in the ACR histogram, and alerts one to use a wider color space.

Here is a plot from Gernot Hoffmann's web site showing the real world surface colors (dotted lines) along with sRGB, aRGB, and ProPhotoRGB gamuts.

I find it generally useful to perform abstract, in practice useless tests - in thought or in real - in order to understand the behaviour and the underlying principles of some phenomenon.

One could say it is useless to shoot brick walls and newspapers instead of looking at the nice picture of the favourite cat (usually a crop in 25%), as proof for the high quality of a lens. I am on the other side. I am shooting abstact and/or uninteresting subjects and peeping the non-demosaiced images, when I judge a lense (the test of bokeh is an exemption).

That's still not relevant. That's only a spatial issue (resolution), not a color issue.



Sure there is. Have you ever shot red or blue LEDs? You can have one color channel almost clipping, while another is down near the noise floor; the green response is weak but significant. I have seen shadows cast from an LED in the "right" color on bulbs, viewed in the "wrong" RAW channel. Looks like they aren't even on at all, and you have to take a shot with just that one bulb on to see it in the "wrong" channel, buried down in the noise. But no one claimed 100% saturation.



Who said two channels? Only one has to mostly miss it for it to be considered saturated. How do you record a yellow or violet or cyan laser? How do you display it? You sometimes must depend on two colors for recording and displaying saturation (not ideal, of course). Saturation is emulated by dropping the weaker channel more than it is really recorded in RAW, while applying a push into the ceiling" curve to the other two colors.

I have a deep blue 77mm filter that drops everything on the red side of green down into the noise. I have a "pad" of gel samples, some of which, when looking at a spectrum through a diffraction grating, look black in certain ranges.



Could you rephrase that, and make your point clearer.

True, but that is not all it depends on. Any comparison discussion about DOF between formats must be concerned with equivalent FOV, or else it's pointless. The question should be - can I essentially capture the same image on two different formats. In the case of extremes of DOF range, the answer is often no.

Sensor format and lens have to be mixed to find out which sensor can provide more bokeh for a given lens.
In more understandable terms: given a lens with a maximum aperture (e.g. a 85mm f1.8) and shooting over the same subject, a 5D will achieve a narrower DoF (i.e. more bokeh) than a 40D, because the 5D will allow us to get closer to the subject to obtain the same FoV over the subject, then reducing DoF.
This is what I meant.

The same story applies in wide angle lenses: wide angles lenses do not distort perspective since perspective does not depend on the lens but on the distance to the subject. But wide angle lenses allow us to get closer to the subject, and it is then when perspective changes.

The ability to abstract details is essential for the understanding of issues.

The spatial distribution is not "only" when evaluating the color, but irrelevant in the current context.


I shot red laser light. The blue was very low, but the green was considerable, about 25% of the red. Of course, this depends on the wavelength.



1. In the RGB model the yellow is a composite color. There is no important difference between yellow and any other color.

2. *All colors* appear in *at least two kinds of pixels* to a meaningful degree. This fact makes it possible to differentiate between colors of the same "range", for example between red colors; otherwise the Bayer sensor would not work with multicolor.

Why are you telling me all this?

Most of what you seem to be saying here is what I have already said, yet you seem to be disagreeing with me.

I have serious difficulty following a conversation with you.

Let's recapitulate it. I posted earlier:



and your opinion was, that this played no role in the phenomenon Jonathan observed:



All this time I tried to explain you, that the de-mosaicing does change pure colors in composite ones, and thus it causes the contradiction between the raw and RGB histograms.

Another issue is, that even with tri-colored pixels a conversion would have to occur with like effect (i.e. converting pure colors in composite ones), but this does not really matter, for the sensor in the experiment is a Bayer one..

Anyway, I stand by my assertion, that the de-mosaicing is the cause of the phenomenon Jonathan observed with the LED lights.

De-mosaicing has nothing to do with it. A Foveon sensor would behave similarly, with a single-wavelength light source such as an LED or laser stimulating a response in more than one channel. Converting pure colors to composite colors is a product of the frequency response of the color filter array, pure and simple. You can observe this in the RAW data before demosaicing occurs:



This histogram is from one of my red LED test photos.

Then you are out on a limb, IMO. The only difference of a bayer CFA here is one of spatial resolution. There is no color-related magic in co-location. The only reason for demosaicing over straight, independent interpolation of each isolated color plane, is that you get to abstract per-pixel luminance. If you wanted a 3MP conversion of a 12MP sensor, you wouldn't even need to demosaic; it would serve no purpose. You would just interpolate each color plane (maintaining its offset), and then downsample or bin to 3 MP.

FWIW, I've posted my 1Ds and 1D-Mark II UniWB RAWS online at:

http://www.visual-vacations.com/images/2008/

I stated just above, that this would occur with tri-colored pixels as well. However, creating composite colors from Bayer pixels is part of the de-mosaicing process.

Guillermo,

I uploaded two pairs of images, both 4 EV apart. They were shot with exposure bracketing, using a wired remote control, 2sec time delay; the tripod was on carpit. One pair shot with MLU, the other with MLU + live view.

http://www.panopeeper.com/Demo/MLU-2EV.CR2
http://www.panopeeper.com/Demo/MLU+2EV.CR2
http://www.panopeeper.com/Demo/LiveView-2EV.CR2
http://www.panopeeper.com/Demo/LiveView+2EV.CR2

The roof of the house can be taken for comparison, the leaves not, because of a slight breeze.

I wonder if you achieved better matching images with MLU than I did.

But the 5D has over 20% more pixels than the 40D. 3570x12.8/10.1=4524.

It's not clear to me what role the extra 2 bits of the 40D plays, but I recall John Sheehy mentioning that they don't appear to be producing any more real levels than 12 bits would produce.

It is not reasonable to reduce this question to the number of bits; it's not so simple.

The 5D creates about ~3570 levels. The 20D creates ~3970 levels. The 40D creates ~12800 levels at ISO 100, and ~15200 levels at higher, full stop ISOs.

This is about 3.5 times more at ISO 100 than that of the 5D. It is open to debate, how many of the 12800 levels of the 40D are really informative, but IMO it is clear, that the 5D is far underequipped with the 3570 levels.

On the practical side: I have no problem with carrying around lots of uninformative levels, they don't disturb me the least; however, the lack of levels in the shadows is an impediment. I would not exchange my 40D for a 5D based on image quality (based on other considerations even less).

Though it would be interesting to see a detailed, pixel-peeping comparison between the 40D and 5D, with the very same lens(es).

Probably no better, and no worse hehe. I will check that anyway.
When I have compared shots put in PS layers, the location of scene's elements was indistinguishable pixel by pixel.

Pano, you are talking about captured levels, but don't forget that interpolated levels are 16-bit on both machines, and photographs are made both from captured and interpolated levels.

Of course, if 40D levels are more precise than 5D's levels (4 times more precise to simplify), 40D's interpolated levels will also be more precise, so we could with no doubt say that 40D's images are more exact in defining the right level values.
But regarding tonal richness, after developing what you will have is:
- 40D: pixels with 1 channel taking 14 bit values, and 2 channels taking 16 bit values.
- 5D: pixels with 1 channel taking 12 bit values, and 2 channels taking 16 bit values.

Tonal precision in 40D is higher, but tonal richness (number of total different levels achieved) can be considered the same on both machines, and thus danger of posterization or banding due to lack of levels.
Moreover in the areas where the increased tonal precision could be really enjoyed (lowest f-stops) noise makes the image unusable so 40D cannot really take advantage of its increased no. of encoding bits.

In fact Leica's M8 RAW files are 8-bit. Yes, they are very cleverly distributed in a non-linear way, but they are 8 bit after all and even in the shadows are les precise (there is more gap between each pair of captured encoded levels) than 12-bit linear. Would you say Leica's M8 enconding produces noticeably lower quality images than any 12-bit linear camera?

I would exchange a 40D for a 5D if my priorities were strong bokeh and wide angle for example. In other words, FF.

Do you mind posting a pair of such shots?

My aim is to achieve raw pixel by pixel coverage; what you see are de-mosaiced pixels.



The number of tones in the resulting image is a non-issue. However, the de-mosaicing can create tones, but it can not create true details. If two pixels of the sensor can not differentiate between two slightly different tones of the subject, then that is lost, no matter how many tones you can fae by the de-mosaicing.



According to DCReview, the DR of the 40D is 9.1 stops in ISO 100 and ISO 200 (IMO this is incorrect, it is somewhat higher in ISO 200), and that of the 5D is 8.2 stops at ISO 100. The noise may be less with the 5D, but I doubt the correctness of those evaluations.

A higher dynamic range means, that the 40D does *need* more levels, and it can utilize them better.

Anyway, I would like to see a DR comparison with Jonathan's mehod, because the DPReview test regards only the noise, not the details.



Conditionally, yes. The M8 produces a crippled raw; it is good for nothing but change the white balance. When you need strong adjustments, the limits of the 8 bit will come forward.

Why do you think Leica's digital back has ***16*** bit depth? Is the digital back 256 times better than the M8?



Forget about the bokeh. No camera will make a lens suitable for good bokeh, no matter of the DoF. I don't know, what "strong" bokeh is, but if you like nice bokeh, you need a suitable lens, and that will not be an F4 lens.

The above statements are contradictory. Tones are a non-issue. Then by implication, an insufficient number of tones will limit rendering of detail. What did you mean to say?

Actually bit depth and resolution are not highly correlated. There is often little difference in resolution between an 8 bit JPEG and a raw file rendered into 16 bits.

Gullermo wrote:

regarding tonal richness, after developing what you will have is:
- 40D: pixels with 1 channel taking 14 bit values, and 2 channels taking 16 bit values.
- 5D: pixels with 1 channel taking 12 bit values, and 2 channels taking 16 bit values

We are talking about two kinds of levels: those sensed and those "created". Already the 256x256x256 tones are plenty in most cases, so the issue is not if the number of combinations from 3x12 bits are enough.

Any number of resulting levels does not help, when the levels of a channel are not enough to distinguish between details.



I did not mention resolution. Different areas of an image may be indistinguishable because of lack of resolution OR because of lack of tonal levels.

An abstract example: half of the scenery consists of something of a constant color, the other half too, but slightly different color. If they are not different enough to be "seen as different" by the sensels, then the two halves appear identical - this has nothing to do with resolution.

Interesting example. Would this be possible to demonstrate visually? I can appreciate that from a purely technical point of view, one could have the two halves of an image a very slightly different hue so that in order to technically (at the machine level) distinguish between them, you might need 14 bit or 16 bit capture instead of 12 bit, but would the eye be able to distinguish between such subtle differences that required this huge increase in the number of levels afforded by 14 or 16 bit processing?

We already know, for example, that the last stop of an ETTR exposure, in 12 bit, produces a far greater number of levels than the eye can detect, and maybe the penultimate stop too.

I cannot Panopeeper, I see clear that the same lens capturing the same scene (same FoV over the subject) will provide at maximum aperture a shorter DoF in the 5D than in the 40D. This is what I call, maybe wrongly, strong bokeh. And it's what I am looking for in FF (apart from wide angle), a higher capability of differentiating the subject from the background/foreground.

I agree with the rest of your post although I think none of the statements will mean a real and noticeably improvement on 40D's images. 9 f-stops is simply not enough to really NEED extra bits, 12-bit cameras have showed to be able to capture with a reasonable good definition 9 f-stops (the Sony A700 for instance).
I haven't done tests over the 5D's DR at ISO100, but I assume it is lower than 40D's, so in the 5D the extra bits are even less necessary.
For a good DR in high contrast scenes, both machines need extra exposures. Any possible advantage of 14 bits is definitively gone in these circumstances.

I will try to post a couple of overlapping shoots tonight. Didn't have time to check yours sorry.

The 5D needs about 2500 levels at ISO 100 and 1090 levels at ISO 1600 for normal (non-stacking) use. That's for the RAW data; for conversion, of course, it is always good to force extra precision if the converter loads RAW data aligned by the LSB.
Any bit depth (or number of levels) that brings the standard deviation of a black frame above 1.4 ADU is inefficient for storage purposes.

The level count can cause low-contrast detail to disappear. Try loading my DR test chart in PS and reducing the number of levels and see how many you can throw out before the smallest, lowest-contrast text starts disappearing.

Such a test would be totally irrelevant, if the decrease in levels weren't in the original RAW data. Conversions deteriorate much more quickly from quantization.

The point of such a test would be to get a rough idea of how many levels per stop are needed to avoid visually apparent banding/posterization and the disappearance of low-contrast fine detail.

But of what? You replied to Ray, who was talking about RAW bit depth, and consequently, RAW levels. Quantizing them is not the same thing as quantizing finished conversions. RAW data is far more quantization-resistant than full RGB images with realistic tone curves.

WTF??? In the highlights, yes, but certainly not in the shadows. Try zeroing out all the lower-order bits of a RAW file so that it is effectively 8-bit, and run it through any RAW converter. Quantization will be much worse than if you convert to 8-bit after conversion especially in the shadows.

It depends on the absolute luminousity of these values as well. The minimum difference between two luminousities (in proportion, not in absolute value) depends on the range these luminousities are in (in absolute values). See the paper James has linked to a few posts above.

Well, "strong bokeh" is not a generally used term, at least I don't know that, so you can use it in whatever sense you want to. However, "nice bokeh" is a well-known term, and even though it is not well defined, most people are in agreement in the judgement of bokehs.

The "quality" of bokeh depends primarily on the lens; it is a special quality of the lens. Larger DoF does not guarantee nicer bokeh. For example the Canon 80-200mm f/2.8L shot in my bokeh collection shows, that that lens is not a "bokeh lens". The 50mm f/1.4 yields a medium good bokeeh only, no matter if at f/1.4 or f/2.8.

The reasons for creating a nice or bad bokeh are often discussed and guessed, but as far as I see it, this is rather shamanry. I read already, that some of the MTF curves indicate the bokeh quality, but I don't see, why. However, the number and shape of the aperture blades appear to be important, except with the maximum aperture, I guess.

All zeros is not the proper way to do it, unless you're also going to move the blackpoint down to compensate. Using "1000" (for 12-bit) is what you need to replace with.



I was thinking more along the lines of 16 vs 14 vs 12 etc. If you go to 8 significant bits, you're going to see the quantization (especially with low ISOs).

With only 2500 levels, quite a few of the 256 RGB levels would be wasted.



I wonder how you calculate this, specifically for the 5D.

If you read The Dialog, the first and largest section of the new Rawnalyze manual, you find how to zero out some bits (chapter Changing the original raw data). However, shifting out bits instead of zeroing is better with Canon cameras, because of the black level (explanation inside).

About 1000 of the ~3590 levels currently used are already wasted by read noise.



"1.4 ADU" is not specific to the 5D.

It is a general figure based upon my experiments with quantizing RAW data to various bit depths, and observing at what read noise level, in ADUs, is the number of levels sufficient. Anything resulting in much less than about 1.4 ADUs starts to get quantized visibly, and anything above 1.4 ADU looks the same as anything around 1.4 ADU.

I'm using 1.4 as a generous safety margin; 1.25 works quite well, as can be seen in the cameras that have 14 bits @ about 5 ADU.

This does not depend on noise, nor on DR. It depends on the transform function and on the target range. With sRGB this is not a real issue, but with Adobe RGB 98 one would need much more (over 10000) levels in order to utilize all 256 target levels (and where is that from 10-bit printers and monitors?).



I know. I asked for a specific example of your calculation, with raw values and noise levels.

I would like to see data supporting this assertion. As Roger Clark points out, noise in DSLR cameras consists mostly of shot noise until you get fairly deeply into the shadows. Here is a noise model taken from Roger's analysis of the 1D MII. I think you overemphasize read noise.

The table shows shot noise and read noise for the 1DMII at ISO 100 with shot and read noise expressed in electrons on the left and DNs on the right. A 12 stop range is covered, but if the darkest f/stop needs 8 levels, the effective DR is limited to 9 stops by bit depth considerations alone, disregarding noise, and as shown by Norman Koren.. The analysis assumes that the camera places the highlights at 4095, but in practice it may be lower as you indicate.

In zones 0 through 7, the noise is predominately shot noise, and these zones are considered shot noise limited. There are 4080 levels. Only zone 8 is read noise lilmited, and it contains 8 levels. If some posterization is acceptable, you could include 7 more levels for a total of 15. Where do you get the value of 1000?

Of course, DR may be limited by noise as well as posterization, but the S:N of 1.49 in the deepest shadows, while quite low, does contain image information.

Define "fairly deep(ly)". Do you realize how far below saturation "middle blue" is in tungsten WB?

Also, if someone wants to make a statement to say current cameras are shot noise limited, it's the kind of thing that can't be proven right or wrong easily, because the point of reference is arbitrary. My question is, why would someone even say that? What purpose does a statemnt like that serve except to be academic fluff or filler?

And, as it turns out, shot noise is not the only noise in highlight area of cameras. You should be seeing Emil Martin's updates to his web pages soon, where he discusses another form of noise present in highlights (usually only visible in the top 1 to 2 stops of the sensor's DR).



Unlike you and Roger, I have actually investigated what pure shot noise would look like in the deep shadows, and it is a relatively beautiful thing, compared to the reality of read noise. Blacks are actually black.



And that is significant because ...? Would anyone really expect SNR limits to occur in the highlights?

Neither the sensor nor the camera are shot noise limited.



I wrote 1000 for ISO 1600, not ISO 100, which I said would be served well by 2500 levels at ISO 100.

(Edit: I thought you were referring to my statement about ~1000 levels being enough. Now that I think of it, you may be talking about the "1000 levels already wasted" - that referred to the fact that about 3500 are used in the 5D at ISO 100, and only about 2500 are needed (not including some room for negative read noise). I do not mean to imply by this that the levels should requantized to this amount; I simply mean that the original capture could have been at this many levels in the original digitization, for a more compressible file size, with infinitessimal loss of signal.)

The basic idea is that if you only need enough linear levels for the read noise to be at least 1.4 ADU, then you can scale down the number of levels by 1.4/readnoise, to get a baseline for the minimum number of levels for digitizing.



Of course it does. You can't improve on it by having more than a certain number of levels, though, as the signal is just an average of lots of noise. Rounding the noise values out to infinitessimally more accurate values with more levels does not help the signal come through the noise.

I have measured this visually, with extreme stretching of the levels, with RAW data from numerous cameras, and simulations, and 1.4 ADU is my safe limit.

OK, here is the same type of analysis showing read and shot noise for the 1DMII at ISO 1600. Again I would like to see your data supporting your assertions, not some blanket statement from high.

You replied about 1 minute too soon. I just edited the post you replied to, as there were two figures of 1000 or close in my post, and I apparently assumed the wrong reference in the post you replied to.



You need to be a bit clearer about what you don't believe. I felt a very vague feeling in your previous reply.

So let me state the core of what I am saying, and then you can object specifically to something or ask for proof. It's as if I am supposed to know what is apparently illogical to you and defend it. Nothing I've written recently is illogical to me, so I can't figure out what you think is questionable.

Here is a summary of what I have been saying:
"In the face of all the analog noises involved in the readout of a sensor, quantization of any practical significance can only occur when the number of linear levels used is such that the blackframe read noise, in those ADUs, falls significantly below 1.4, for single-exposure RAW images."

As I've stated previously, 1.4 is a conservative value. We can get away with read noise ADUs as low as 1.1 without incident (and that implies even less levels needed).

The tools that I am using to look at these matters are horrible, in terms of the workflow involved in compositing images for comparisons. You can verify what I say for yourself, quite easily, though. Just open two instances of IRIS, and in one, select a RAW from a camera whose read noise is known at the ISO. Then, calculate the division needed to bring that read noise down to 1.4 ADU. Set the threshold sliders to some window down in the shadows, crop an area of interest, and then save out the .fit file. In the second instance of IRIS, load the crop. In this second instance, divide the image by the factor needed to scale the read noise down to 1.4 ADU. Multiply back by that number again (or rescale the threshold sliders, if that is more convenient). What you will see, is exactly the same thing, as far as your eyes can tell, in both instances. Now, reload the crop into the second instance of IRIS, and divide by the read noise in original ADUs (to make the new noise 1.0 ADU). *NOW*, you can see a little bit of quantization. Try again, bringing the read noise down to 0.8, and 0.7 ADU and things fall apart very rapidly. This is true regardless of what level of read noise you started at; you only start running out of useful levels when you get down to 1.4 ADU of noise.

Now, try similar things with highlights. Find the brightest area in a smooth, OOF gradient, and measure its sigma. Then, do the same as before, to bring this down to 1.4 ADU. The bright area that you got the shot noise deviation from will look like it lost no smoothness, but the darker areas that might be in the image will show quantization. Try again with 1.0, 0.8, 0.7, etc. Same principal keeps applying.

It is easiest to do this with single color channels, because they are easier to crop. You can verify, however that the same principle applies to color by carefully cropping so that the RGB CFA patterns in the crops are unaltered, or if you have plenty of RAM, just don't crop at all and just window-in the areas to compare (quantization must be before color conversion, though).

One day you will realize that noise is a hard ruler of appreciable levels, and all the anecdotes about levels and levels per stop and such is usually irrelevant in today's noisy digital photography. All those anecdotes come from noiseless, synthetic graphics and are totally meaningless in digital photography.

I am not telepathic. What assertion do you believe I am making, which needs to be proven?

In any event, I don't see the relevance of the chart you posted to the topic at hand. We're talking about necessary levels in digitization, are we not?

That 1000 levels are lost to read noise.




The chart clearly shows the relative contributions of read and shot noise for the exposure zones, and shot noise predominates in the green area of the chart, which comprises 4080 levels, whereas read noise dominates only in the very deep shadows, comprising 15 levels at most. How can you lose 1000 levels to read noise, when it predominates in only 15?