Nikon D7000 Dynamic Range

[ErikKaffehr]

Ray,

My first experiments with HDR were on my Alpha 100. That camera had less DR than what I have to day. HDR programs were less well developed at that time. So I tested HDR and did not like the results.

Regarding your D7000, I just hope that you see the advantages you expect! Please share the experience.

I normally don't print that large. A2 is my standard size, anything larger is printed on Durst Lambda at a pro lab. I never did any large print from images having excessive dynamic range, so I have not that experience.

Best regards
Erik

Which DSLR are you referring to, Erik?


The question is, '"Will I, or will I not, see the following 2-stop difference in shadows and lower mid-tones, when I compare the new camera with my old Canon 5D (old but full frame)?"

[PierreVandevenne]

Let's suppose that in order to achieve that average charge of 32e- in each pixel in the 11th stop in Camera B, we need to let in 8x the amount of light. Instead of an average of 32 photons per pixel, we need an average of 256 photons per pixel. Instead of an exposure of 1/200th at F11 for camera A, we need an exposure of 1/25th at F11 for Camera B.

Now, it seems reasonable to me, in my ignorance, to surmise or deduce that Camera B will exhibit less shot noise in the RAW file, than Camera A, although Camera A, being a mythically perfect camera, will possibly have lower noise in total because the only noise it will produce (or reproduce) is shot noise.

Camera B on the other hand will likely produce images with noise which is predominantly read noise, thermal noise, fixed pattern noise and the whole gamut of various noise types. Its lower shot noise (in proportion to signal) will not help much in these circumstances.

You might be right, but I'm not convinced yet    .

If you push your example to the extreme, you'll see that a camera with 0.001% effective QE would be even better than your camera B.

Try to look at "shot noise" it from the side of received/measured/counted photons.

[Ray]

If you push your example to the extreme, you'll see that a camera with 0.001% effective QE would be even better than your camera B.

Hi Pierre,
My point is not that Camera B is better than Camera A, but that the signal in Camera B contains a lower proportion of shot noise. If the effective QE of Camera B were 0.001% then a normal exposure in bright daylight, at base ISO, might take 10 seconds or more, instead of 1/100th of a second at the same aperture with Camera A.

In such a situation, it seems to me, that shot noise in Camera B would become totally irrelevant, and noise from sources other than readout noise would increase, such as thermal noise.

Try to look at "shot noise" it from the side of received/measured/counted photons.

Okay! I'll try. But I'm not Einstein and it's very difficult. I have virtually no understanding of why the processing of received light, once it has entered the camera and the shutter has closed, should increase the proportion of shot noise in the final signal, expressed as an electron voltage.

This is how I imagine the shot noise occurs. For the sake of simplicity, let's imagine we have a shutter than opens and closes intantly, faster than the speed of light, and that the entire signal for the entire scene exists momentarily inside the camera body, before it gets processed, and before it passes through the various obstacles on its way to the photodiodes.

Let's quantify the number of photons in that entire signal as being 1 million, for the sake of simplicity. The shot noise should be sqrt of 1 million = 1,000 photons, or 0.1% of the total signal. If that's true, then it's an insignificant amount of noise. Right?

Let's now consider the effects of just one aspect in the processing chain, but the real 'bady', the Color Filter Array which absorbs about half the light that has entered the camera.

In one fell swoop we are suddenly reduced to 500,000 photons. The other 500,000 have got absorbed by the CFA.

Now you (and Roy) seem to be implying that the CFA has a bias towards 'misbehaved, naughty' photons (those that arrived early or late), and lets pass a greater proportion of the 'naughty' photons.

If that's the case, please explain why. The shot noise in a million photons is 0.1% of the signal, whereas the shot noise in 500,000 photons is 0.14%. Why the bias towards 'noisy', 'misbehaved' photons?

[Ray]

Regarding your D7000, I just hope that you see the advantages you expect! Please share the experience.

Thanks Erik. Will try to do. I'm very curious as to why some people give little or no credence to the DXO results.

I tend to think the reasons are either due to personal bias or a lack of understanding of the meaning of the graphs and the differences between the normalised print results and the 'screen' results. The graphs are quite explicit and very accurate in my opinion.

I find it quite uncanny that a year or more before DXO published its test results, I had already determined that the maximum DR of my Canon 5D, from the perspective of discernment any detail, however degraded, was in the order of 11 to 12 stops, depending on the accuracy of the ETTR exposure in the first stop, which I assess by a 2/3rds of a stop negative EC adjustment in ACR, but which the late Gabor (Panopeeper) assessed slightly differently. He reckoned my ETTR exposure was in fact slightly clipped.

DXOMark describe the DR of the Canon 5D as being 11.13 EV. Here are my test results, from about 2 years ago, using Jonathan Wienke's DR Test Target (thanks Jonathan! ).

[Ray]

The image of the 12th stop doesn't show on my browser.

I'll try again.

[Ray]

Oops! Looks like all that noise is producing a larger jpeg. Here's a downsized image.

[ErikKaffehr]

Ray,

I don't know. In my view the data is both interesting and useful. Two issues I may have are:

Don't really like bunching all data in one figure of merit.

The DR measurement is feasible but not very relevant due to the definition used. Nothing wrong with the definition, but it may be less relevant to photography.

I got a mail from Mark Dubovoy, who sometimes seem to be critical of DxO data, saying that the data itself is excellent but is used incorrectly by many posters. This may also be my opinion.

There is a wealth of data in those measurements!

Regarding DR, it has seldom been a problem for me, but I will revisit the issue, and also check with Jonathan's test targets.

Best regards
Erik

Thanks Erik. Will try to do. I'm very curious as to why some people give little or no credence to the DXO results.

[Ray]

For those who are interested, here's the 10th stop (that's 10 stops including the first stop). I think you will agree that this quality could be useful for espionage purposes, determining car number plates in the dark for example.

The big question for me is, will the D7000 be able to turn such an exposure into a photographically useful part of an image in the shadows.

In true Aussie fashion, I'm now taking bets on this issue. I can receive betting payments through PayPal, and to check out the odds I'm offering, visit my website at......... just joking!   .

[Ray]

Regarding DR, it has seldom been a problem for me, but I will revisit the issue, and also check with Jonathan's test targets.

Erik,
It has seldom been a problem for you because you instinctively avoid photographing scenes which you know will not turn out well, considering the DR limitations of the camera. An extreme example, but also relevant for less extreme situations, is the shot of a scene out of a living room window, offering a view of lovely clouds, perhaps a river view, or even a view of the Himalayas.

Without fill-flash, which tends to produce unnatural black shadows unless one has expertly arranged the lighting as a professional might with his truck-load of gear, one ends up with a beautifully exposed view through the window, but the interior of the living room, which might be quite stylish and deserve attention in its own right, is totally and unacceptably noisy.

In such extreme situations, even the D7000 will be found to be lacking, and bracketing exposure is the only solution. But for less extreme situations, the D7000 might do the job without the need for bracketing.

That's the point I'm trying to get across.

[PierreVandevenne]

Now you (and Roy) seem to be implying that the CFA has a bias towards 'misbehaved, naughty' photons (those that arrived early or late), and lets pass a greater proportion of the 'naughty' photons.

No, not really. Essentially, and in this case, you are sampling the signal. There are no naughty photons, no well behaved photons.
If the average stream of measured photons is 1 per exposure, there is a 36.8% chance that your exposure will record nothing, a 36.8% chance that it will record 1 photon, 18.4% it will detect 2, 6.1% that it will detect 3, 1.5% that it will detect 4, etc... If the average stream is 10 per exposure, you'll get 9 / 10 in 12.5% of the cases, more than 11% for 8 / 11, around 10% for 12, apporx 9.5% for 7 etc... (the Poisson distribution is a bit skewed but tends to become quite close to a Gaussian distribution after 20-30 samples). It's not how many photons there were in the first place that matter (if that was the case, there would be close to zero noise in the shot of the light from a distant star eventhough the light was filtered by light years of dust, the atmosphere, etc...) but what we sample out of the photon stream. Saying "here's the shot noise" is saying "there's the uncertainty on the signal we measure". The more individual events in the sample, the closer it will statistically be from the average stream, which we consider the "real value". To go back to the coins example, the 100 toss sample will be on average closer to 50-50 than two individual 10 samples. The 500000 photons at the diaphragm are not your sample, they 150.000 at the sensor are.
Now, for the funny aspect of the issue, if our perceptions were extremely fast samples of the physical world, much faster than the speed at which photons arrive (obviously impossible but still fun to think about it), a shot with 0 photon could be the way I see and remember a scene - my reality - and the same shot with 2 photons could be the way you see and remember it - a different reality...

[Ray]

No, not really. Essentially, and in this case, you are sampling the signal. There are no naughty photons, no well behaved photons.

Of course. I realise this. I was speaking metaphorically and being a bit facetious. 

As I understand this issue of shot noise, in layman's terminology since I'm clearly not a quantum physicist, in any given exposure of a supposed constant light source (say a star as viewed from the atmosphere-free moon), the number of photons that pass through the open shutter of a camera will vary with each exposure, despite everything else (f stop and shutter speed etc) being the same.

Just like tossing a coin will produce a result close to 50% heads and 50% tales after many events, the more exposures we take, the closer the variation in photon count will be to the sqrt of the average number of photons in each exposure. Is this correct?

One interesting feature of Photoshop Extended is the facility to stack a number of images and get one resulting image which consists of the best exposed pixels from each of the stacked images. I haven't pushed this to its limits yet, due to lack of computer power. My experiments as I recall, when I first got CS3E, have consisted of about 6 stacked  images.

The improvement in noise with just 6 images was quite dramatic. I would describe it as a 2 stop difference at high ISO. In other words, 6 stacked images at ISO 3200 after processing, look as clean and sharp as a single shot taken at ISO 800.

Now at high ISO shot noise will be greater, but also SNR in general will be worse.

I have no idea to what extent this 2-stop improvement is due to countering the effects of shot noise, or due to countering the random effects of normal signal processing. Perhaps you can tell me.
   

If the average stream of measured photons is 1 per exposure, there is a 36.8% chance that your exposure will record nothing, a 36.8% chance that it will record 1 photon, 18.4% it will detect 2, 6.1% that it will detect 3, 1.5% that it will detect 4, etc...

If you had a perfect Bayer type Color Filter Array in which the red filter was perfectly transparent to photons of the red wave length, and perfectly opaque to photons in the green part of the spectrum, a single photon of red wave length, after passing through the shutter, would stand only a 25% chance of getting through the CFA, for a start.

With a 'real-world 'CFA in a 'real-world' camera, the chances would be considerably less than 25%, and the chances of that photon knocking an electron off the silicon substrate of the photodiode, would be yet again less, in my humble opinion. I'd give it no chance. 

[hjulenissen]

DXOMark describe the DR of the Canon 5D as being 11.13 EV. Here are my test results, from about 2 years ago, using Jonathan Wienke's DR Test Target (thanks Jonathan! ).

Thank you.

What do you think about this way of measuring DR for highly nonlinear devices, i.e. film? It seems to me that if 20 stops of dynamic range is captured (using e.g. HDR), representing that in a globally tonemapped 8-bit jpeg file as "20 stops of DR" is somewhat misleading, even if some details can be discerned that far apart?

-h

[PierreVandevenne]

source (say a star as viewed from the atmosphere-free moon), the number of photons that pass through the open shutter of a camera will vary with each exposure, despite everything else (f stop and shutter speed etc) being the same.

Just like tossing a coin will produce a result close to 50% heads and 50% tales after many events, the more exposures we take, the closer the variation in photon count will be to the sqrt of the average number of photons in each exposure. Is this correct?

On the whole that is the idea.

[quote}
One interesting feature of Photoshop Extended is the facility to stack a number of images and get one resulting image which consists of the best exposed pixels from each of the stacked images. I haven't pushed this to its limits yet, due to lack of computer power. My experiments as I recall, when I first got CS3E, have consisted of about 6 stacked  images.

The improvement in noise with just 6 images was quite dramatic. I would describe it as a 2 stop difference at high ISO. In other words, 6 stacked images at ISO 3200 after processing, look as clean and sharp as a single shot taken at ISO 800.

Now at high ISO shot noise will be greater, but also SNR in general will be worse.

I have no idea to what extent this 2-stop improvement is due to countering the effects of shot noise, or due to countering the random effects of normal signal processing. Perhaps you can tell me.
 [/quote]

Again roughly speaking, stacking images improves the number of photons collected and therefore increases the S/N ratio. If you want to improve the signal to noise ratio by a factor of 4, you need 16 exposures. Assuming your exposure period doesn't overflow your electron bins, a single exposure equivalent to the total exposure of the stacked exposure should have a bit less noise than the stack because the individual components of the stack each suffer from read noise.

In practice, there are many sources of noise that should be considered.

For example, if you have a camera with a very low read noise,  stacks won't lose that much quality compared to a single equivalent exposure. If you have a sensor that overheats easily, a stack of short exposures is likely to be better than a single longer exposure because it will suffer less from dark current. etc... etc...

The rule of thumb for amateur astronomy for example is to use the camera at its "optimal ISO" and collect as much photons as possible. That's basically attemtpting to obtain the ISO-less camera ejmartin wants.

[bjanes]

The image of the 12th stop doesn't show on my browser.

I'll try again.

Ray,

Your image has strong pattern noise in the deep shadows, typical for many older Canon's. If you read Emil Martinec's excellent treatise on noise and dynamic range, he classifies that as one type of PRNU (pixel response non-uniformity). Such pattern noise has a more detrimental effect on image quality than a mere standard deviation would indicate. BTW, it is typically greater for CMOS than CCD, since the former has an amplifier for each pixel.

Furthermore, such noise would not be included in the engineering definition of DR, since full well (in electrons) and read noise are determined by subtracting duplicate frames to eliminate fixed pattern noise (see Roger Clark).

DXO does not describe their DR measurement in detail, but they use a circular array of neutral density optical filters somewhat analogous to the use of a Stouffer wedge (where the steps are arranged in a linear fashion). DR is taken as the maximum signal divided by the step where the SNR (signal to noise ratio is unity). This would include pattern noise and veiling flare.

If readers think that this DR is too liberal, one can use the expanded DXO SNR graphs as explained by Emil.

Regards,

Bill

[BJL]

Ray,

    given your offensive and irrelevant accusations about my opinion of DXo, I probably should not bother to reply at all. After all, my comments carefully avoided any judgement on the accuracy of DXo's measurement: my main point n recent posts is simply that your possible evidence of dark noise affecting IQ with one camera (the 20D) is irrelevant to whether dark noise can significantly effect IQ in another camera with far less dark noise (the D7000).

But I am replying to help you with a factual misunderstanding: shot noise depends on the number of photons detected, and is roughly the square root of the number detected, limiting SNR to the square rot of the number of photons detected. So as long as your hypothetical less sensitive camera B is using a higher exposure level (longer exposure time and or lower f-stop) and so receiving more photons in order to detect the same number of photons, and produce the same number photo-electrons, it will have the same shot noise, not less. Otherwise, ND filters would solve all out shot noise problems when long enough exposures can be used!

For the same reason, a lower light level (cloudy day) does not reduce the effect of shot noise on IQ in any useful way: in reducing both signal and shot noise, less light received/detected gives a worse ratio of signal to shot noise, along with a worse still ratio of signal to dark noise, and so an overall worse ratio of signal to noise.

Hopefully I can leave it to to real experts like Bill Janes and Emil Martinec from here in.

[bjanes]

Hopefully I can leave it to to real experts like Bill Janes and Emil Martinec from here in.

BJL,

Thanks for the compliment, but I am hardly in the same league as Emil, who is a real scientist (professor of physics at a major research university). I am merely a medical doctor (pathology and internal medicine) who finds the science of photography to be quite interesting. I find your posts to be very instructive and accurate and would include you as an expert (I gather that your professional background is in engineering).

Regards,

Bill

[Ray]

Ray,

    given your offensive and irrelevant accusations about my opinion of DXo, I probably should not bother to reply at all. After all, my comments carefully avoided any judgement on the accuracy of DXo's measurement: my main point n recent posts is simply that your possible evidence of dark noise affecting IQ with one camera (the 20D) is irrelevant to whether dark noise can significantly effect IQ in another camera with far less dark noise (the D7000).

Don't be so sensitive, BJL. I was merely using my own test results of the 20D in conjunction with DXO results, in order to predict, or get an impression of, the sort of noise levels I could expect with the D7000. I was also offended by your bringing irrelevant distractions to the issue, such as the 20D being old technology. The age of the 20D does not appear on the DXO graphs, nor on my test images.

I am predicting that the lowest 4 stops within the DR range of the 20D, at base ISO and at the pixel level, will be cleaner in the D7000, and they will be cleaner to the extent which can be observed when those same 4 stops have been captured using the 20D at ISO 1600, with the same exposure.

But I am replying to help you with a factual misunderstanding: shot noise depends on the number of photons detected, and is roughly the square root of the number detected, limiting SNR to the square rot of the number of photons detected. So as long as your hypothetical less sensitive camera B is using a higher exposure level (longer exposure time and or lower f-stop) and so receiving more photons in order to detect the same number of photons, and produce the same number photo-electrons, it will have the same shot noise, not less. Otherwise, ND filters would solve all out shot noise problems when long enough exposures can be used!

Fair enough! I stand corrected (provisionally). The ND filter is a good analogy, but I don't understand why such increase in exposure to the light before it enters the lens should not reduce the proportion of shot noise in the detected signal to some degree.

However, I can understand why such a reduction in shot noise might never be noticed. If it were possible to reduce shot noise by use of an ND filter and appropriately longer exposure (and I'm certainly not asserting that it is. I'm just trying to wrap my head around this issue), such modest reductions in shot noise would be swamped by the increased 'electronic' noise of the longer exposure required. This is not something that I would know how to test, using my own camera equipment and ND filter. It would be a waste of time trying.

For the same reason, a lower light level (cloudy day) does not reduce the effect of shot noise on IQ in any useful way: in reducing both signal and shot noise, less light received/detected gives a worse ratio of signal to shot noise, along with a worse still ratio of signal to dark noise, and so an overall worse ratio of signal to noise.

There's a misunderstanding here. I failed to see the relevance of this analogy and merely pointed out to Roy, who made the analogy, that a 'correct' exposure of a scene on a dull day would have less shot noise as a proportion of total signal, than would a 'correct' exposure of the same scene on a sunny day.

[Ray]

Thank you.

What do you think about this way of measuring DR for highly nonlinear devices, i.e. film? It seems to me that if 20 stops of dynamic range is captured (using e.g. HDR), representing that in a globally tonemapped 8-bit jpeg file as "20 stops of DR" is somewhat misleading, even if some details can be discerned that far apart?

-h

Hi,
I would never attempt to do these sorts of comparisons using film. For a start, you have the DR specifications of the scanner to take into consideration. I've found quite often when scanning film that it has been necessary to make at least two scans, one for the highlights and one for the shadows, and then merge to HDR.

I dont see the compression of a 20 stop DR into an 8 bit jpeg, or onto a print with its 6 or so stops of DR, as being misleading. When the eye peruses a scene with a 20 EV brightness range, it cannot perceive that entire brightness range in one fixed stare, just as it cannot perceive the entire field of view within one fixed stare, excluding peripheral vision. To encompass and register the entire scene in the brain, the eye has to shift its gaze, up and down, or from side to side, and the eyes' pupils dilate and contract according to the brightness of that portion of the scene one is viewing at any specific time.

Compressing the entire DR range into 6 stops on a print allows one to appreciate the whole scene in one glance, unless the print is very large or very wide and one is very close to it.

[Ray]

Ray,

Your image has strong pattern noise in the deep shadows, typical for many older Canon's.

Don't I know it!    This pattern noise was the first thing I noticed when I received my 5D about 5 years ago. I thought the camera might be faulty and returned it. The second camera seemed to have slightly less pattern noise in the shadows, so I kept it. But by that time, I had got a clear impression from reports on the internet that this limitation of the 5D was normal.

I recall John Sheehy had a few suggestions of ways to reduce the pattern noise, but I never bothered to explore such techniques. There's a limit to the amount of time I'm prepared to spend in order to extract the Nth degree of detail from an image. I've got far too many images to process and too little time  .

[hjulenissen]

Hi,
I would never attempt to do these sorts of comparisons using film. For a start, you have the DR specifications of the scanner to take into consideration. I've found quite often when scanning film that it has been necessary to make at least two scans, one for the highlights and one for the shadows, and then merge to HDR.

I dont see the compression of a 20 stop DR into an 8 bit jpeg, or onto a print with its 6 or so stops of DR, as being misleading. When the eye peruses a scene with a 20 EV brightness range, it cannot perceive that entire brightness range in one fixed stare, just as it cannot perceive the entire field of view within one fixed stare, excluding peripheral vision. To encompass and register the entire scene in the brain, the eye has to shift its gaze, up and down, or from side to side, and the eyes' pupils dilate and contract according to the brightness of that portion of the scene one is viewing at any specific time.

Compressing the entire DR range into 6 stops on a print allows one to appreciate the whole scene in one glance, unless the print is very large or very wide and one is very close to it.

That is assuming that the non-linearity is close to the desired one. If it is not, then one will have to apply a new non-linearity to get the desired response. I would guess that one major reason why people wants a large DR is because they assume that this will give them a "large space to work within", extracting and presenting different parts of the tonal range as they see fit.

If the DR is wrapped in a heavily quantized/noisy delivery format, doing significant non-linear adjustement may not be that easy.

"Optimal" tonemapping seems to be closely supported by theory, just like you suggest, yet many photographers are strongly against it. Using whatever nonlinear, non-controllable mapping inherit in analog film seems to be favored amongs many photographers. Why is it so? Only history/habits, or some technical reason that I have missed?

-h