Real World Dynamic Range

[Nick Rains]

I'm looking for some enlightenment by the more technically minded LL folks.

Here's the thing; MFDBs and good DSLRs are generally accepted as having a DR in the 11-13 stops range, depending on how you measure it and so on.

My Leica S2 produces a 16bit file which obviously has the potential for 16 stops but in reality it's much less than this - I understand that much. However, using the theoretical 16bit range it would follow that the top one stop contains 65536 levels and at the 7th stop down, 512 levels.

Now, here's my problem.

I just bought a ND2.1 filter, that's 7 stops. It's a Formatt brand from the UK and seems quite neutral so I'm happy. The meter shows it being 6 stops different but it's close enough for the purposes of this discussion.

So, I thought I'd test it today and shot a ColourChecker Passport at an exposure (without the ND) which gave me an almost perfect exposure where the white patches measured around 250 in ACR. It was 1/500 second at f2.8. I put on the filter and shot the same thing at the same exposure to get a 7 stop underexposed image.

I figured that if the camera has a DR of, let's say, 12 stops, then a 7 stop underexposed image should still capture plenty of tones to examine.

The results were not at all like I expected. The image was almost utterly black with the whitest patch barely registering on my Eizo monitor (it reads about 0,0,5 in PSCS5, 16bit file).

So, if that's 7 stops, how come the camera is supposed to have a 12 stop DR? In this situation it looks more like 6.

What am I missing here?

[Sheldon N]

I think that most discussions of practical DR involve shooting a RAW file where the range of the scene actually contains 9 or 10 stops of dynamic range. This could be a scene like shooting from an interior looking out a window to a sunlit area.  You set the exposure to just avoid clipping the highlights, or even slightly overexpose them but only to the point where they can be pulled back from a RAW file and still keep acceptable quality/detail. Then look down into the shadows (which will appear black) and attempt to use exposure compensation or a tone curve to bring detail up out of the darkness and see how far down you can go. It takes a spot meter to do this because you need to see what the actual brightness of each area of the photo was at the time you took the photo.

Another good practical way to test this is to use a full spectrum light source (ie flash or daylight) and a scene with both light/white textured items and dark toned items with detail in them. Then shoot a series of overexposed through underexposed images in one stop increments. Play around with the files in your RAW editor of choice, see how far you can push the highlights and how far you can pull the shadows without crossing that subjective line of whether it would make a good final image for you. You can then calculate what the total practical dynamic range for your purposes is. I'm guessing it's probably closer to 9 stops in the real world, depending on where your personal quality threshold is.

A color chart or step wedge isn't really ideal for this kind of thing, because it only contains tones without detail. You want to be able to see the results on a real subject. It's difficult to put a firm number on dynamic range, because it involves the issue of how you process a RAW file and in what software, and making a decision about how much noise in the shadows is too much noise. Basically it's a subjective call about how far you can overexpose and still pull quality information from the highlights combined with how far you can underexpose and still pull quality detail from the shadows without noise becoming too objectionable.

My understanding of the whole 12 bit vs 14 bit vs 16 bit issue is that it only relates to how finely/accurately the camera can separate the tones between one another, not describing how much shadow detail or highlight headroom there is. Think of it as the difference between the number of steps on a staircase as compared to the overall length of the staircase.

[Nick Rains]

I'll try that - Thanks.

I just tried a different test - a shot of my office which has some opaque glass windows. The difference in light levels between the bright windows and a dark space under a table measures 9 stops according to my spot meter. I set the exposure to place the windows at +2.5 to just clip them.

On the 5D2 using 100 Fill in LR3 I got lots of noise but visible detail in the darkest parts, and on the S2 with 50 Fill I got a quite decent result. That seems to show two things:

1. The S2 is way better in the shadows than the 5D2 (Duh)
2. Underexposing by 7 stops using an ND filter is in some ways different to an image with a wide brightness range. Not sure what the difference is though.

[Nick Rains]

Cinema cameras are tested with transparent stepper charts. So in a lab test method it works quite well to show what range you get.

A simple example, Zacuto tested the footage from a 5DII to Kodak and Fuji motion films. They used a trans stepper which showed the 5DII had about 8-10 stops (don't remember the specifics) and motion film showed considerably more.

David Muench has made amazing images for decades on large format trans film with maybe 5-6 stops of range. I say make some art and worry less about the tech, it's better for the soul.

LOL.

Yes indeed.

This all stemmed from working on a new workshop talk about RAW and bit depth etc. I realized I was touting the party line and had never actually done the legwork myself. Having the ND2.1 filter arrive spurred me into action.
 

[Nick Rains]

I think I have a handle on this now - I was comparing my ND image to a normally exposed shot when I should have been seeing how far I could over expose it too, before full detail loss. it's nothing to do with the ND, I just didn't extend UP the range, only down.

Forgetting the ND, if I shoot my Colour Checker using a wide range of full shutter speed steps on the S2 and then look at the sequence of images I can see details in about 12 or 13 different images. That's not to say each is correct, or can even be corrected, but I see some detail, even if it's almost washed out through over exposure, or almost, but not quite, totally black.

This now agrees far better with the 12-13 stop DR quoted by many, esp the manufacturers.

[LKaven]

My understanding of the whole 12 bit vs 14 bit vs 16 bit issue is that it only relates to how finely/accurately the camera can separate the tones between one another, not describing how much shadow detail or highlight headroom there is. Think of it as the difference between the number of steps on a staircase as compared to the overall length of the staircase.

Since the sensor is linear to the best approximation, the RAW numbers are also linear, and one bit ~= one stop.

So the stairs are fixed, until some later stage in the processing where profiling and gamma and white balance and so on come into play.  [If the length of the staircase were fixed somehow, how would the model accommodate variations in dynamic range and what would the length of the staircase be fixed at and by what convention?] 

The reason this confuses people is that the maximum signal level is fixed as a reference level up and down the signal chain.  The extra dynamic range is manifest in the low order bits.  This gives the illusion that the dynamic range is fixed.  Most monitors only use 6-8 bits anyway, though high dynamic range monitors capable of 16 stops of resolution are in development.

[bradleygibson]

Nick,

This area is very complex, partly because once you're looking at a presentable image on-screen all kinds of changes have been made to your original, raw pixel values.

It is non-trivial to back out these changes that every raw processor makes to clear and simple intensity values.

No tool that I've ever come across is simple if you want to look at the raw bits, but you want to dig in a bit further, take a look at dcraw.  It's is a great (and free) way to pull out the raw bits as a TIFF file.  The TIFF file won't be much to look at, but load it up into Photoshop and you'll be able to see the pixel values just as your camera recorded them before your raw converter adjusts for black level/noise floor, channel gain, linearization of your camera's ADC's, gamma, color space, etc. etc. before viewing on-screen.  (It's no wonder the values you were measuring didn't reveal a simple relationship to exposure!)

Ideally, in the linear domain, double the value equals double the light (photons).  The variation of numbers of the number of stops of dynamic range you hear usually come from different people's interpretation of where to start and stop measuring this range.  I do like DxO Mark's definitions for dynamic range, and while there are other valid ways one could define a useful dynamic range, being able to compare many systems across a single (reasonable) standard is a Good Thing in my book.

Hope that helps,
-Brad

[Christoph C. Feldhaim]

Honestly - I don't understand at all, why bit level and DR should in any way be related.
It simply doesn't make any sense.

You can have 4 bits bit depth and 16 stops dynamic range in the same moment.
It would only be a quite coarse mapping of the input range (light intensities) to the output range (numbers).

Bit depth and dynamic range are in no way related.
But, of course,  you need enough bit depth to have a fine enough mapping for your purpose and to avoid banding.

[sojournerphoto]

"Honestly - I don't understand at all, why bit level and DR should in any way be related.
It simply doesn't make any sense.

You can have 4 bits bit depth and 16 stops dynamic range in the same moment.
It would only be a quite coarse mapping of the input range (light intensities) to the output range (numbers).

Bit depth and dynamic range are in no way related.
But, of course,  you need enough bit depth to have a fine enough mapping for your purpose and to avoid banding."


This is a mistake that has been trumpeted around the forums for too long and by too many people who ought to know better - not referring to Christoph.

Sensors count photons and the output voltage from the sensel is proportional to the number of photons detected - a linear relationship. The voltage is converted into a binary number by the analogue digital converter. This is set to (theoretically - some aren't at some isos) reach it's maximum count at base iso when the sensel has a voltage equivalent to the most photons it can count. The number of bits then determines the number of lower voltages it can count and so each extra bit doubles the number of possible counts (it's binary) and halves the minimum voltage that can be measured - again it's a linear measure. Half the voltage is half the photons is an extra stop of dynamic range.

In practice, noise gets in the way so the camera can't resolve the lowest voltages cleanly and at high isos the count would be more than one per photon, so the full adc based dynamic range is not available.

Mike

[Nick Rains]

Thanks Mike, that makes sense - I was about to agree with Christopher but this clears it up. You are saying that, in the absence of noise, a 16 bit sensor can always record 16 stops (doublings) of light intensities - by definition. The usable DR is then limited by signal noise levels.

Does this sound about right?

[Christoph C. Feldhaim]

Hmm .. I am not sure if I can yet agree, though I'm no tech expert.

If you add 1 bit to an existing sampling scheme the amount of possible discreet values is doubled.

You would use this to make the steps that can be sampled smaller and thus add an additional f-stop at the lower end of exposure, where only few photons are recorded.
The full-well capacity of the sensor would not be changed.

I'd agree, if you were to change the bit depth in a non noise-limited sensor, a perfect sensor.
But since most sensors have a noise to signal ratio which doesn't allow to fully use more of these photons at the lower end, adding a bit in an already well resolving sampling scheme would most likely just record deeper into the noise.

So - practically and if I'm not completely mistaken (which might be so and then please again correct me) in the end the noise to signal ratio determines the usable DR, and not the mathematical resolution / bit depth of the A/D converter. The banding problem of course is not touched by this.

[mrenters]

I think I have a handle on this now - I was comparing my ND image to a normally exposed shot when I should have been seeing how far I could over expose it too, before full detail loss. it's nothing to do with the ND, I just didn't extend UP the range, only down.

Forgetting the ND, if I shoot my Colour Checker using a wide range of full shutter speed steps on the S2 and then look at the sequence of images I can see details in about 12 or 13 different images. That's not to say each is correct, or can even be corrected, but I see some detail, even if it's almost washed out through over exposure, or almost, but not quite, totally black.

This now agrees far better with the 12-13 stop DR quoted by many, esp the manufacturers.

I did a similar experiment to yours a couple of months back, comparing a P65+ to a Canon 5DII.  I used a B+W 10 stop ND filter and photographed a Color Checker and a stepped gray chart. I then used Capture One to process both images and adjusted the Exposure and Highlight/Shadow recovery sliders to see just how much detail was recoverable. I talked to Jeff Schewe about it and he suggested trying it again with some objects that had texture.

Martin

[Graeme Nattress]

Often raw converters use a black level somewhat higher that the sensor's actual black level. This can hide noise near black, but can also drastically reduce dynamic range in a test when you're gaining up a very dark image.

Also, unless you have access to the raw data code values, it can be hard to determine if you're actually clipping or near clipping or not. The tonal development curve in some raw software will intentionally over-expose bright regions, not least to hide funky colour artifacts from differential clipping of the RGB channels.

So - your 250 on the colour picker (98%, 95% when you un-do gamma) could easily be a stop or more below clipping in the raw data.

Graeme

[NikoJorj]

I figured that if the camera has a DR of, let's say, 12 stops, then a 7 stop underexposed image should still capture plenty of tones to examine.
The results were not at all like I expected. The image was almost utterly black with the whitest patch barely registering on my Eizo monitor (it reads about 0,0,5 in PSCS5, 16bit file).

You tested the dynamic range of an entire system : camera, AND (standard) treatment, AND display.
To see what the camera alone can achieve, you'll have to equal things on the treatment and display side, ie to try to make both images look the same on the display. If you can do it, then you're within camera DR.

Forgetting the ND, if I shoot my Colour Checker using a wide range of full shutter speed steps [...] and then look at the sequence of images I can see details in about 12 or 13 different images.

That seems more correct ; try to correct each image (brighten the underexposed ones, and vice-versa) to see where there is still enough detail and tones. As said, textures can help to see when noise starts to drown image content on the underexposure side ; on the other side, try to see when false or washed colors begin to show (note that the raw converter used still plays a role : some handle noise better than others, some others handle overexposure better, etc...).
DR is the difference between these two limits.

I think that most discussions of practical DR involve shooting a RAW file where the range of the scene actually contains 9 or 10 stops of dynamic range. This could be a scene like shooting from an interior looking out a window to a sunlit area.  You set the exposure to just avoid clipping the highlights, or even slightly overexpose them but only to the point where they can be pulled back from a RAW file and still keep acceptable quality/detail. Then look down into the shadows (which will appear black) and attempt to use exposure compensation or a tone curve to bring detail up out of the darkness and see how far down you can go.

You won't get (exactly) the same results with that method, because having the highlights and shadows in the same scene means that the lens will produce some veiling glare, depending on the place and quantity of highlights (and depending of the lens and aperture used of course). This has the double consequence to raise the shadows (remember pre-exposure in the darkroom?), making them easier to record, and to flatten their contrast, making them less legible.
On one hand, it's much more of a real-world experiment, because DR always matter in a unique scene in practice, and I'd thought that with real-world lenses veiling glare can often be the limiting factor.
On the other hand, it throws many more variables in the equation and make results much less repeatable.

[bjanes]

Honestly - I don't understand at all, why bit level and DR should in any way be related.
It simply doesn't make any sense.

You can have 4 bits bit depth and 16 stops dynamic range in the same moment.
It would only be a quite coarse mapping of the input range (light intensities) to the output range (numbers).

Bit depth and dynamic range are in no way related.
But, of course,  you need enough bit depth to have a fine enough mapping for your purpose and to avoid banding.

It depends on how you use the bits. With current digital cameras the sensor is linear and linear encoding is used. The output of the sensor is proportional to the luminance in the scene. With a bit depth of 4, the minimum luminance is than be recorded is 1 and the maximum is 2^4 = 16, which gives a DR of 16:1. Expressed in f/stops the DR is Log base2 of 16 = 4 stops. Using binary recording simplifies the math.

Sensors that have a logarithmic output are widely used in telecommunications. If we assume a log base 10 is used, the minimum output from the sensor would be 1 and the maximum would still be 16. The luminance ratio for these outputs would be 10^16 : 1. You would have a very high DR but very poor gradation. Log output can be used for high DR recording (see Greg Ward). Another approach is to use floating point representation, which allows the steps to be of variable size.

[Graeme Nattress]

Floating point encoding is just integer encoding with a log curve though, when it comes down to it.

[ondebanks]

The tonal development curve in some raw software will intentionally over-expose bright regions, not least to hide funky colour artifacts from differential clipping of the RGB channels.

Graeme has mentioned an important point. To do the DR test right, one must only look at one colour (R, G, or B) within the Bayer matrix. I use dcraw and imagemagick to pull out the colour planes into FITS format, and IRAF to analyse them properly - no footling around with non-scientific formats or software.

Ray

[Graeme Nattress]

Yes - unless you're intimately aware with the raw software and what it's doing, you have to get access to the raw data to know exactly what's going on. Even then, raw converters are generally built as tools to make images, rather than tools to take measurements. You can certainly use a raw tool to get an indication of what's going on, but it's not going to be the definitive answer.

Graeme

[sojournerphoto]

Thanks Mike, that makes sense - I was about to agree with Christopher but this clears it up. You are saying that, in the absence of noise, a 16 bit sensor can always record 16 stops (doublings) of light intensities - by definition. The usable DR is then limited by signal noise levels.

Does this sound about right?

Yes, that's correct...

up to the point that the count is 1 count per photon detected. At higher isos (gain) the count will be more than 1 per photon so DR falls by one stop per stop increase in iso gain - for example if the well capacity is around 60,000 (very approximately 2^16=65536) electrons then a 14 bit ADC will count 1 for every 4 electrons stored at 'base iso. If the gain is turned up to increase iso by 2 stops then the count will be one for every electron stored (unity gain). If the gain is turned up further then the count will be more than 1 for each electron and so it is not possible to benefit from finer gradation. These numbers are there or thereabouts for a Canon 1Ds3 at 100 or 400iso and above.

Noise is obviously the fly inthe ointment.

Mike

[sojournerphoto]

If you add 1 bit to an existing sampling scheme the amount of possible discreet values is doubled.

You would use this to make the steps that can be sampled smaller and thus add an additional f-stop at the lower end of exposure, where only few photons are recorded.
The full-well capacity of the sensor would not be changed.

I'd agree, if you were to change the bit depth in a non noise-limited sensor, a perfect sensor.
But since most sensors have a noise to signal ratio which doesn't allow to fully use more of these photons at the lower end, adding a bit in an already well resolving sampling scheme would most likely just record deeper into the noise.

So - practically and if I'm not completely mistaken (which might be so and then please again correct me) in the end the noise to signal ratio determines the usable DR, and not the mathematical resolution / bit depth of the A/D converter. The banding problem of course is not touched by this.

Exactly. Also, the signal itself is noisy - search for shot noise:)

Mike