Too much dynamic range?

[nightbreath]

An interesting thought came to me before I went to bed. Below you'll find an assumption that came suddenly to my head, so please don't take it too seriously.

So... Let's assume there are two cameras with similar color tones reproduction abilities, but with different possible lightness level capturing ability. For example:
- sensor of camera A has 12 stops of DR, 16 billion tones it can distinguish
- sensor of camera B has 10 stops of DR, 16 billion tones it can distinguish

Having a flat scene (i.e. low DR scene) on a shot we'll push an image with, say, 8 DR to be captured with both sensors. And then both images will be edited in post to retrieve lacking contrast. So we need to add:
- 4 stops for 12-stop camera
- 2 stops for 10-stop camera

So my point is: with lower DR camera we'll have lower tone delta (difference of the initial color tone in the scene with reproduced tone by the sensor) when processing the low DR shot made using lower DR sensor. That happens because of decreased amount of modifications made to the file to achieve required result.

What do you guys think about that?

[Hector1970]

I was a bit bamboozled with the question.
Having read it I wondering with the camera with the higher dynamic range it could have if the dynamic range of the scene was wider gone two more stops but because its not required the extra dynamic range would be redundant therefore the cameras are equal for that scene.
The conclusions being if you own the lower dynamic range camera don't take high contrast scene or else you should have bought the camera with the higher dynamic range (which is probably more expensive) . This would lead to severe buyers regret and a need to sell the camera purchased at a loss. This would be sad

[Edwin Herdman]

So my point is: with lower DR camera we'll have lower tone delta (difference of the initial color tone in the scene with reproduced tone by the sensor) when processing the low DR shot made using lower DR sensor.

I understand the question.

It's similar to the Adobe RGB versus sRGB space question - at least to hear Ken Rockwell tell it, aRGB sacrifices tonal gradations for gamut.  sRGB should allow for finer gradations in color change.

I think that before sensors will reach the limit of the current color space in terms of usable bit depth, the recorded bit depth will be increased.  If I am up to date with my reading, the analog-digital converter (ADC) is often claimed to be the stumbling block in this, but it is probably really the sensors themselves.  The ADC has to provide more precision than the original sensor capture data in order to preserve quality, so the actual bit depth ends up being a bit arbitrary (a best tradeoff).

[Policar]

It's a valid concern, but someone did some posterization tests and found that noise is still the limit by a very big margin. Even 8 bit JPEGs are good enough for most purposes.

I have worked with Alexa footage (14 stops DR quoted, but in practice it feels like dramatically more than any dSLR) and it's compressed into a 10 bit wrapper... almost no posterization no matter how you grade it. I wouldn't worry.

[tim]

So my point is: with lower DR camera we'll have lower tone delta

It doesn't work like that.  Suppose you have a sensor with a dynamic range of 10 stops.  That's (roughly) equivalent to saying that the noise is 1 part in 1000 of the maximum signal (2^10 = 1024), so you can only distinguish 1000 different shades.  Even if you put a 16-bit ADC on the sensor, you're still limited to 1000 shades.

Now I've glossed over a lot of things in the paragraph above.  For one thing noise is not constant over the tonal range: it gets larger at higher signal amplitudes, so in the example you'll actually be able to distinguish less than 1000 shades.

There's also a reason why you might use an ADC with a larger dynamic range than your sensor.  If you used a 10-bit ADC on a sensor capable of distinguishing 1000 shades, every shade in a scene would be assigned a definite value in the photograph.  Subtle gradients in the scene would be rendered as stepwise increments in the photograph, and under heavy processing this could become visible as a posterisation-type effect.  But with a higher bit-depth ADC the steps would be blurred out by the electronics noise so that they are not visible.  There's no extra information in the photo, but it looks nicer.

[helpful]

This question is totally relevant, and mathematically justified. A camera's sensor at the hardware level is analog. If it records a total dynamic range of 14 stops (let's say) and this data is converted to a 14-bit digital signal, then one could say that 2^14 = 16,384 tones are being recorded per channel, ignoring noise.

If the camera's sensor records a total dynamic range of 10 stops (let's say), and this data is also converted to a 14-bit digital signal, then there are still 16,384 tones being recorded per color channel, ignoring noise.

However, those tones are representing a range of bright to dark which has 4 stops (16 times) less variation in tone. Ignoring noise, tones are recorded with 4 stops (16 times) as much sensitivity to tiny shifts in color/contrast. This is true only for tones within that limit of 10 stops of dynamic range. Tones outside that range are lost, which is the drawback.

A camera sensor X could be designed with the same amount of noise as a camera sensor Y, but a lower dynamic range for X and a higher dynamic range for Y. If the signal were accurately converted to the same 14-bit RAW digital output, then under the given assumption that both sensors had the same amount of noise, then there would always be better gradation between tones in the output from camera sensor X, but better resistance to blown highlights / lost shadows in the output from camera sensor Y.

Nothing in the world can increase without a trade-off, and that is definitely true for dynamic range as well. If all other factors are held constant, increasing dynamic range has the drawback of decreasing gradation in tone. In the limiting case of infinite dynamic range, then all levels of signal would be rendered as a single flat tone, just like a delta "spike" function in Fourier analysis corresponds to a signal of all wavelengths.

[tim]

No really, it doesn't work like that.  The sensor noise determines both the dynamic range and the number of tones which can be distinguished.  They're inextricably linked.  If you want better accuracy of tones you need to reduce the noise, which automatically increases the dynamic range.

[tim]

Here's a simplified example.  Suppose you have two sensors:

Sensor-A gives a signal between 0 and 1000mV, and has a noise of 10mV.
Sensor-B gives a signal between 0 and 1000mV, and has a noise of 1mV.

Sensor-A can just distinguish signals corresponding to 500mV and 510mV, but it cannot distinguish signals corresponding to 500mV and 501mV.  In other words Sensor-A can distinguish 100 levels.  And the ratio between the maximum and (average) minimum signals is 100, which is the same as saying it has a dynamic range of 100 (= 6.6 stops).

Sensor-B meanwhile can distinguish 1000 levels and has a dynamic range of 1000 (= 10 stops).

The noise determines both the dynamic range and the precision of the intermediate gradations.

[Sporgon]

Interesting to read what helpful has to say. Is this why the tonal graduation of the 5dmkiii and 1dx is so good ? Many pictures from these cameras have a "film" like quality which looks to me to be the way the chip is handling graduation to highlight and low light. Photographers that I know who have the Mkiii have really noticed this: it is superior to the mkii. That's why I'm so surprised in reading the criticisms directed at the Mkiii/1DX vs the D800. (from what I have seen so far ) the D800 does not have this film like quality, and surely the best of film is the Holy Grail of digital?

[jukka]

Show me one pictures taken with 5dmk2 and mk3 and the difference you are talking about (i have them)
Show me one picture from 5dmk3 and d800 and the difference you are talking about  (I have also d800)

[jukka]

No really, it doesn't work like that.  The sensor noise determines both the dynamic range and the number of tones which can be distinguished.  They're inextricably linked.  If you want better accuracy of tones you need to reduce the noise, which automatically increases the dynamic range.

and this  is Canons big problem with their old read out circuits , the read out noise from 5dmk3 (as one  example) are 12 times higher than d800 at base iso, therefore 5dmk3 has about 11 stops DR and Nikon 14 stop

[nightbreath]

I didn't want to focus only on DR (possible lightness levels) in my initial post, my main idea was to draw attention to color tones reproduction abilities of a sensor.

When there are X stops of DR available, this is lightness range captured, i.e. not matter if each individual pixel has purple / yellow / gray tone. So idea of the topic is to ask everyone whether "there's life outside DR".

As I see it: DR is longitude (or Y coordinate), color tone is latitude (or X in two-dimension representation).

The more gradation steps there are in each dimension the better a sensor is.

So the questions are:
- Should we look at both dimensions instead of referencing DR only?
- Does the initial question about contrast improvement work? (contrast amplification affects lightness level and color tones at the same time, so initial colors can be thrown away when editing an image)
- Why is there difference in color tones reproduction of 1D vs. 5D lines? What do we miss?
- Is that noise at the floor so important? Or there are other things that take part in the game?

[nightbreath]

I'm waiting for someone smart to chime in and describe whether my initial thought can affect real life shooting, or whether there's something important we don't pay much attention to. One of things that may be related to the topic and was confirmed by several photographers is:

difference in color tones reproduction of 1D vs. 5D lines

Even if you take 1D Mark IV, the pixels are better suited to color modification than those from 5D Mark III. Similar thing is mentioned in 1D X review here:

One odd thing I’ve noticed before, now definitely in this test- there’s a big difference in how the 5D series and 1D series interprets shadows. The 5D2 notoriously expresses shadow detail with a purple hue and the 5D3 shows this trend still continues; whereas the 1D4 has to be pushed to its limits before the purple shows up in the shadows. I thought this was mostly due to the crop sensor excluding the lens edges, but the 1Dx- full frame- follows the trend of the 1D4 by not degrading shadows with purple hues. I’ve asked a Canon rep why this is and he’s forwarded it on up to a uber-geeky tech, so hopefully we’ll get an explanation about this. It certainly stands to reason that color-integrity in the shadows is a perk of paying for the higher model, but I’d still like to know what’s the difference. I’ll let you know if we get an answer.

[jukka]

1D series has more expensive electronic chain and also when it comes to shielding etc and 1d series is probably also  better matched in terms of RGB
The old 1dsmk3 has a better response regarding middle tones than 5d mk2 mk3 series an can be seen in a even colored surface.
There also different CFA  in  the old 5d  compared to 5dmk2 mk3 and some experiencing the colors better in the old 5d
Canon changed their color filters  (not so dense ) in order to gain more light/ increasing sensitivity

[helpful]

One extremely theoretical way to view dynamic range is the ratio of the sensor's noise level (in photons / quantized energy units) to the sensor's white point (in the same units).

This extremely theoretical way of viewing noise and dynamic range to be equivalent is useless in the real world.

Viewing dynamic range as a number of stops that can be represented in an image, as I did, is much more practical and less theoretical.

The engineering truth is that noise exists also at each point between the black level and the white level of the sensor.

Consider a camera whose sensor is exposed the 10 brightness levels of Ansel Adam's zone system, in proportion to the camera's actual level of dynamic range (arbitrarily scaled to begin with 25 "units" for level 0):

Level 0 = black level = less than or equal to 25 units of true light, which is lost within 25 units of random noise
Level 1 = 50 units of true light energy +/- a different amount of random noise, which is varies considerably from sensor to sensor and at different levels--it is not necessarily 25.
Level 2 = 100 units of true light energy +/- an even different amount of random noise, etc.
...
Level 10 = 25,600 units of true light energy +- zero noise because at this point the sensor site is fully saturated and is desensitized to any further amplification of signal as well as any noise

Clearly there is fluctuation going on in the interval between level zero and level 10. The number of gradations in tone that can actually be distinguished by the camera depends on the integral/summation of the sizes of the variable noise through all brightness levels, to determine an average noise level, and then dividing the white point energy level by that average noise level that separates distinguishable levels of gradation in tone.

To say that dynamic range and the number of gradations in tone are equivalent, or to say that either one of them can be determined simply from the ratio between the black noise level and the white level, is as stupid as defining people's adult heights merely by their weight at birth.