"Using the default JPEG settings the Mark III delivers somewhere in the region of 8.6 stops of dynamic range from ISO 200 to 800, dropping slightly at ISO 1600 and by more than a stop at ISO 3200."
I have measured 9 stops on my 1Ds3 and 9.6 stops on my 1Ds2, using the same measurement technique as outlined earlier by Pondria. Others have corroborated my findings for these two cameras.
Ben Horne will probably chime in and say "yes but the 1Ds3 has far less noise in the shadows so the DR is actually more usable". I would not disagree with that, but if you define DR limits as clipping of shadows / highlights, the 1Ds3 has less than the 1Ds2.
Also, my 40D came in at 9.6 as well, but let me assure you that I care for the 1Ds3 image "feel" much more. Numbers just can't describe everything.
The 1Ds2 has more latitude when it comes to highlights - mostly what Stan said. More details can be lifted out the 1Ds3 shadows with less risk of noise being an issue. So I guess there should be much net difference, but I do feel the 1Ds2 was easier to manage in high DR conditions (at ISO 100). I think Canon believed that we would all be running around with Highlight Tone Priority engaged and patted themselves on the back for a job well done. I'd rather have a sensor that captures the DR without the added in-camera ISO tricks.
I found my 1Ds3 files to have a much greater DR than my 1Ds2. I'll agree with the others that shadow noise was very well controlled on my 1Ds3 as well, didn't notice much difference on the highlights, but, using LR and the "recovery tool", it was apparent that there was much more data in there than my 5D.
It is mathematically impossible for a jpeg to have more than 8 stops of DR values because it is made up only of 8-bit numbers. Anything more than 8 stops has to be a mathematical or measurement aberration. Admittedly those numbers can represent a wider DR but only by assigning more than one bit of real life brightness to every bit in the numeric value. And then you lose the fine gradation between adjacent tones.
Even with raw shooting (using 12- or 14-bit numbers) several bits are just lost in "pure noise". I don't believe that the camera only loses one stop at ISO 3200 compared with ISO 100. They all lose far more than that. Noise may be well managed at ISO 400 but DR is already falling and is much lower at ISO 1600. Also, ISO 3200 is an internal fudge in which an underexposed ISO 1600 image is taken and brightened up, and that alone must lose a whole stop.
As noise levels rise the effective or acceptable DR (it is somewhat subjective) must fall. Noise levels rise as ISO rises. You can filter much of that noise at but because it already buried the fine shadow details you will also lose those details when filtering, and hence lose effective DR.
The important things are that better cameras produce better DR at any specific ISO, but that all of the cameras produce significantly less DR at higher ISOs than at lower ISOs.
Alan321 wrote:
It is mathematically impossible for a jpeg to have more than 8 stops of DR values because it is made up only of 8-bit numbers. Anything more than 8 stops has to be a mathematical or measurement aberration. Admittedly those numbers can represent a wider DR but only by assigning more than one bit of real life brightness to every bit in the numeric value. And then you lose the fine gradation between adjacent tones.
Even with raw shooting (using 12- or 14-bit numbers) several bits are just lost in "pure noise". I don't believe that the camera only loses one stop at ISO 3200 compared with ISO 100. They all lose far more than that. Noise may be well managed at ISO 400 but DR is already falling and is much lower at ISO 1600. Also, ISO 3200 is an internal fudge in which an underexposed ISO 1600 image is taken and brightened up, and that alone must lose a whole stop.
As noise levels rise the effective or acceptable DR (it is somewhat subjective) must fall. Noise levels rise as ISO rises. You can filter much of that noise at but because it already buried the fine shadow details you will also lose those details when filtering, and hence lose effective DR.
The important things are that better cameras produce better DR at any specific ISO, but that all of the cameras produce significantly less DR at higher ISOs than at lower ISOs.
Nope, this is incorrect. It doesnt matter how many bits are used to store the data, its all about the light that is registered as 0 (black) and MAX (white). This could be done with 4 bits or 400 bits! A 14 bit RAW doesnt capture more DR that a 12-bit RAW it just has more in between data. It's like having a cake and splitting it into 12 or 14 pieces; the cake is the same but you can have more or less pieces depending on how its cut.
Allan Bruce wrote:
Nope, this is incorrect. It doesnt matter how many bits are used to store the data, its all about the light that is registered as 0 (black) and MAX (white). This could be done with 4 bits or 400 bits! A 14 bit RAW doesnt capture more DR that a 12-bit RAW it just has more in between data. It's like having a cake and splitting it into 12 or 14 pieces; the cake is the same but you can have more or less pieces depending on how its cut.
Allan
Just to clarify, lets put some numbers to this to help explain. Lets say absolute black is recorded on the sensor at level 0 and absolute white at 100. Now an 8-bit format could save each value as:
0, 14, 29, 43, 57, 71, 86, 100
whereas a 14-bit format could save as:
0, 8, 15, 23, 31, 38, 46, 54, 62, 69, 77, 84, 92, 100
Both formats can still display the 0 and 100 therefore they have the same dynamic range but the 14-bit one has a smoother gradient from 0 to 100.
In practical terms, I can say that I don't seem to notice any less DR on my 1Ds3 than on my 5D's. I can also verify what Ben says, that 14 bit raw files allow you to rescue a whole grip of info from the shadows, so the trick is to expose like reversal and avoid clipping the highlights. There's little need to expose to the right.
Allan Bruce wrote:
Nope, this is incorrect. It doesnt matter how many bits are used to store the data, its all about the light that is registered as 0 (black) and MAX (white). This could be done with 4 bits or 400 bits! A 14 bit RAW doesnt capture more DR that a 12-bit RAW it just has more in between data. It's like having a cake and splitting it into 12 or 14 pieces; the cake is the same but you can have more or less pieces depending on how its cut.
I do expose to the right (since I shoot in low stage light and often the highlights are just an errant light or two) and find that I can recover up to 1.5 stops of over exposure in RAW-not JPEG.
Alan321 wrote:
It is mathematically impossible for a jpeg to have more than 8 stops of DR values because it is made up only of 8-bit numbers. Anything more than 8 stops has to be a mathematical or measurement aberration. Admittedly those numbers can represent a wider DR but only by assigning more than one bit of real life brightness to every bit in the numeric value. And then you lose the fine gradation between adjacent tones.
This is incorrect. Bit depth limits DR only for linear encoding of data. The nonlinear gamma transformation expands shadows and compresses highlights, allowing the jpeg to encode more than 8 stops of DR.
Even with raw shooting (using 12- or 14-bit numbers) several bits are just lost in "pure noise". I don't believe that the camera only loses one stop at ISO 3200 compared with ISO 100. They all lose far more than that. Noise may be well managed at ISO 400 but DR is already falling and is much lower at ISO 1600. Also, ISO 3200 is an internal fudge in which an underexposed ISO 1600 image is taken and brightened up, and that alone must lose a whole stop.
True. ISO 1600 loses four stops of highlights relative to ISO 100 on an absolute exposure scale, but gains back a little over two of those at the shadow end due to lowered read noise relative to the photon signal. Each stop of ISO loses a stop of capture at the upper end as the signal is amplified past the digitization range of the ADC; typically Canons get back those stops at the shadow end up to ISO 400; ISO 800 somewhat less, very little from ISO 800 to 1600, and essentially nothing beyond that.
Allan Bruce wrote:
Nope, this is incorrect. It doesnt matter how many bits are used to store the data, its all about the light that is registered as 0 (black) and MAX (white). This could be done with 4 bits or 400 bits! A 14 bit RAW doesnt capture more DR that a 12-bit RAW it just has more in between data. It's like having a cake and splitting it into 12 or 14 pieces; the cake is the same but you can have more or less pieces depending on how its cut.
Allan
Nope, this is incorrect ;-)
Dynamic range is set by the noise level at zero signal relative to the signal level at saturation. The latter sets the highest recordable signal, the former sets the lower bound of tonal gradation that can be distinguished from random noise fluctuations. A 14-bit raw does not encode finer tonal gradations than a 12-bit raw when the noise is larger than the quantization interval between 12-bit data (which it is in current Canons). You can't distinguish arbitrarily fine tonal gradations when the noise makes the tones jump around randomly by a minimum amount set by the noise in the electronics.
Dynamic range is set by the noise level at zero signal relative to the signal level at saturation. The latter sets the highest recordable signal, the former sets the lower bound of tonal gradation that can be distinguished from random noise fluctuations. A 14-bit raw does not encode finer tonal gradations than a 12-bit raw when the noise is larger than the quantization interval between 12-bit data (which it is in current Canons). You can't distinguish arbitrarily fine tonal gradations when the noise makes the tones jump around randomly by a minimum amount set by the noise in the electronics....Show more →
So yes, in the absence of noise, increased bit depth means smoother tonal gradations. Now consider the effect of noise, adding gaussian noise of width 12 8-bit levels in photoshop; here is the 8-bit gradient:
The presence of noise before digitization of the signal obliterates the distinction between lower and higher bit depth, so long as the noise exceeds the distance between levels of the lower bit depth digitization (thus avoiding posterization).
While this example was done with 8-bit vs 5-bit so that it could be visualized on an 8-bit monitor, the same principles hold for 12-bit vs 14-bit raw data. The noise in Canon cameras exceeds 4 14-bit levels, which is the distance between levels in 12-bit data (2 fewer bits means four times the level spacing), so there is little practical utility to the extra two bits. (There is of course a difference between earlier 12-bit cameras and newer 14-bit cameras in terms of image quality; it has essentially nothing to do with the increased bit depth, and everything to do with other technological improvements in the newer cameras.)
So yes, in the absence of noise, increased bit depth means smoother tonal gradations. Now consider the effect of noise, adding gaussian noise of width 12 8-bit levels in photoshop; here is the 8-bit gradient:
The presence of noise before digitization of the signal obliterates the distinction between lower and higher bit depth, so long as the noise exceeds the distance between levels of the lower bit depth digitization (thus avoiding posterization).
While this example was done with 8-bit vs 5-bit so that it could be visualized on an 8-bit monitor, the same principles hold for 12-bit vs 14-bit raw data. The noise in Canon cameras exceeds 4 14-bit levels, which is the distance between levels in 12-bit data (2 fewer bits means four times the level spacing), so there is little practical utility to the extra two bits. (There is of course a difference between earlier 12-bit cameras and newer 14-bit cameras in terms of image quality; it has essentially nothing to do with the increased bit depth, and everything to do with other technological improvements in the newer cameras.)...Show more →
Very well presented technical argument. Currently 12-bit is sufficient to digitise the dynamic range of the analog signal recorded by todays sensors. If sensors can achieve DR above 12-bit by lowering the noise floor, then higher bit rate A/D conversion will be required.
Allan Bruce wrote:
Just to clarify, lets put some numbers to this to help explain. Lets say absolute black is recorded on the sensor at level 0 and absolute white at 100. Now an 8-bit format could save each value as:
0, 14, 29, 43, 57, 71, 86, 100
whereas a 14-bit format could save as:
0, 8, 15, 23, 31, 38, 46, 54, 62, 69, 77, 84, 92, 100
Both formats can still display the 0 and 100 therefore they have the same dynamic range but the 14-bit one has a smoother gradient from 0 to 100.
Allan
Just to clarify the relevance of noise (described very well above) I will modify your example. If the sensor is capable of measuring a high DR analogue signal, then you need high bit-depth to digitise that signal. Currently we are limited by sensor technology.
Dynamic range of the analogue signal is the ratio of the maximum signal (saturation point of the photosites) over the lowest recordable signal above the noise floor.
For ease of explanation I describe the maximum value of the analogue signal as an arbitrary value of 128 (using a power of 2 as it makes things easier).
Firstly, lets assume that the lowest recordable analogue signal above noise =1. The dynamic range = 128/1, which is 7 bits or 7 f stops. You will have no problem digitising this signal with an 8-bit A/D converter.
Now lets assume the lowest recordable signal above noise floor is 0.25. DR = 128/0.25 = 512 = 9 bit = 9 f stops. If you had an 8-bit A/D converter you would not be able to digitise the full 9 stops of DR because 8-bit A/D conveter can only record a maximum value of 256.
Conclusion: Bit depth of the A/D will limit the dynamic range of the digital image if it is less than the dynamic range of the sensor.
Above you stated that using use a 4 bit A/D converter doesn't effect dynamic range. It does!!!
We all work with digitised images, not the analogue data recorded at sensor. A 4 bit A/D conversion will severely clip the dynamic range that a sensor is capable of measuring!!
You can confirm by posterising one of your images to 16 levels in Photoshop.
