dcmiller wrote:
There's more tonality below middle gray. It's more than just the darkest shadows. With wide dynamic range shots don't block up as much. But I hate for reality to get in the way of math-based internet conjecture. Carry on.
As in : "That may work in practice, but I'm not so sure about theoretically "
EltonTeng wrote:
So basically the 14-bit files are worthless unless we work in a 16-bit pp environment? Am I incorrect?
What Kamil says is true, but I will say that if you are an SLR-point and shooter, you will see very little if any advantage. What I mean by that is if you primarily shoot JPG, and leave the camera settings pretty much at the default...you probably won't see any significant difference.
I think I can see it in images with strong color especially blacks/reds.
Then again memorex always fooled me!
I think this is one of those features to reserve judgement on until this camera is really out there and all the convertors have been tuned to squeeze every drop out of the RAW images.
until dpreview publishes their detailed review with their standardized method of dynamic range testing, that we can compare to the 5D, 1DII, 1DsII etc...
TP120-31: Transmission step wedge with 31 steps of 1/3 stop each, 10 stops total, 6x6 size. Lay on a light table, photograph with a macro lens, check where the shadows disappear. (Photo is just an example, the only picture they provide is a 21-step 1/2 stop wedge).
I bought one of these for my 5D and it was very illuminating as to usable dynamic range with different contrast curves, post processing settings, highlight recovery etc...
Another misunderstanding I see - the extra two bits helps ALL tonal range, not just shadow area. It's like, you have two more digits after the decimal point. Instead of 100, 10, 1, you have 100.xx, 10.yy, 1.zz. Everyone gets two more digits.
But my argument is that the extra two digits won't help if they are under the noise level.
Pondria wrote:
Another misunderstanding I see - the extra two bits helps ALL tonal range, not just shadow area. It's like, you have two more digits after the decimal point. Instead of 100, 10, 1, you have 100.xx, 10.yy, 1.zz. Everyone gets two more digits.
But my argument is that the extra two digits won't help if they are under the noise level.
I think that what you say is true for linear scales. One extra digit, gives 10 times better description to every value. I think this is not the case here. expose to the right article As I understand this, lower stops will have the less benefit from the extra bits.. or the same level of benefit, but not the same detail level to begin with.
Pondria wrote:
Another misunderstanding I see - the extra two bits helps ALL tonal range, not just shadow area. It's like, you have two more digits after the decimal point. Instead of 100, 10, 1, you have 100.xx, 10.yy, 1.zz. Everyone gets two more digits.
But my argument is that the extra two digits won't help if they are under the noise level.
In most human sensory perception the difference between 100.1 and 100.2 is much much smaller (imperceptable) than the difference between 0.1 and 0.2. So if we are adding bits to increase resolution rather than expanding top end dynamic range then the extra digits will not end up helping humans except for the low values.
until dpreview publishes their detailed review with their standardized method of dynamic range testing, that we can compare to the 5D, 1DII, 1DsII etc...
Hmm...
DP Review are not sufficiently clear about their technique for the validity of their dynamic range data to be asessed. If they take their data from .jpg files or a .tif with a colour depth of 8 bits per channel, all they are measuring is the combined effect of the tone curve and the intrinsic limitation of the file format.
If they are working from .tifs with 16-bits per channel they are getting closer, but the results will still be skewed by the raw conversion software which - in my experience - clips at both ends [perhaps deliberately to avoid converting noise or non-linear highlights].
A raw file with 12 bits per channel should in principle take 2048 samples from the brightest stop, 1024 from the next brightest then 512, 256 and so on down to a single sample from the twelfth stop. Of course, one sample per stop is pictorially useless so we should really only count the stops down to where the sampling error is, say 5% which will give just over seven stops of useful dynamic range: Provided that we have not run into noise by that level and provided that the camer and raw converter don't clip the highlights.
In principle [if not quite in practice] 14 bits per channel would take 8192 samples from the brightest stop then 4096, 2048 and so on for successive stops. If you stopped counting where the sampling error is 5% you now have nine stops of dynamic range to work with - provided that you are not limited by noise first,
When converting that into a visible image you would still apply a tone curve to select the same range of values - regardless of the dynamic range.
The benefit of the high dynalic range would be that you would have more headroom [or exposure latitude if you like] for exposure [unless you still followed the conventional advice of exposing to the right] and much more processing headroom if you wanted to work with a more limited part of the tonal range.
If the on-camera histogram was set so that it did not go right to the highlight end of the data which was actually recorded - say it stopped one stop short - that would give you a whole stop of over-exposure latitude at no practical cost to the colour depth of the tones you wanted to capture.
I don't see how 14 versus 12 bits would change dynamic range that appreciably unless the sensor itself is changed. The extra bit depth will certainly improve noise, but how will it possibly affect the sensor's inherent sensitivity?
Light doesn't fall on the sensor in stops. A stop is only a relative measure. The most common measurement of light intensity in photography is cd/m^2 (candelas per square meter).
If a sensor is not electronically responsive to light below a certain cd/m^2, or if it completely blows its signal to light above a certain cd/m^2, then it's not going to produce any meaningful signal after the A : D conversion. The limit on the sensor happens before the A : D conversion.
If a sensor can relay a meaningful digital signal over a 2^10 range of light intensity (10 stops), then giving it 14, 18, 24 bits is still not going to make it any more inherently sensitive to very low light levels or very high light levels.
So I think the greater bit depth is going to marginally expand detail fidelity in shadows, where detail already exists but it's very noisy. I simply don't see how the bit depth of the digital space has anything to do with the sensor's inherent analog sensitivity -- because it doesn't.
As mentioned above the editing latitude for the captured stops will change, because you've captured analog signal with greater precision. You've encoded it such that small changes won't lead to gross rounding errors. But that's not a change in dynamic range -- it's just a change in latitude.
Think of it this way -- your digital bathroom scale can probably measure weight from 10 up to 350 pounds, more or less. You can't weigh out milligrams of sugar on it, nor can you weigh the tonnage of cargo on a freighter. If you upgrade the scale so it can now report an extra 5 significant figures, all you'll change is the precision, not the range -- so you'll get 10.00000 up to 350.00000 pounds. But the limits on its sensitivity don't change.
I want everything the sensor can give me and not have part of it thrown away by a limited ADC! ADCs are CHEAP! so give me 24 bits like the Pentax and let ME chose which bits I wish to throw away. If you don't digitize the low bits the info is not there. As for noise in the low bits, thats my choice of how I process it.With good DSP you can extract infomation at least 6 Db below the noise floor.
The fundamental limit is the signal to noise ratio of the sensor: So even if an extra two bits gives you an extra two in the shadows for the same quantization error, its difficult to believe that Canon have improved the signal to noise ratio by two stops [6 Db].
So: In practice you might be able to deliberately underexpose by a stop or even two to give youself over-exposure latitude with less of a penalty in the shadows - BUT only if you expose raw images and commit yourself to correcting for the underexposure before conversion or produce 16-bit per channel output.
For a correctly exposed image you are still going to be most concerned about the top six stops and eventually convert the image to 8bits oper channel for printing or display. A 12 bit raw image already provides more headroom than you would ever need for heavy processing [save for some Microphotographic or astronomical applications].
The practical cvalue of 14 bit raw files is very limited.
Paul Gardner wrote:
I want everything the sensor can give me and not have part of it thrown away by a limited ADC! ADCs are CHEAP! so give me 24 bits like the Pentax and let ME chose which bits I wish to throw away. If you don't digitize the low bits the info is not there. As for noise in the low bits, thats my choice of how I process it.With good DSP you can extract infomation at least 6 Db below the noise floor.
...the pentax that throws away that 24 bit info and gives you the same 12-bit RAW that everyone else gives you? pure marketing BS, that pentax "24-bit capture"
I fear that's where we're headed here, too. just look at scanners if you want to see a terrible example of manufacturers saying "they want 12, now 14, now 16 bits? sure, give it to them, they won't even know that the last 5-8 bits are all useless noise"
fourfa wrote:
I fear that's where we're headed here, too. just look at scanners if you want to see a terrible example of manufacturers saying "they want 12, now 14, now 16 bits? sure, give it to them, they won't even know that the last 5-8 bits are all useless noise"
Yeah, scanners are a good example. Scan in 48-bit mode and all you get is a bigger file. Drum scanners can achieve exceptional resolution and with a noise frequency that is far below the frequency of film grain, so they really are delivering more detail; but an Epson 4490 or something is not delivering much more to a 16-bit file than to an 8-bit file.
There's little point in a huge bit depth unless your sensor is actually delivering more information into it.
The fundamental limit is the signal to noise ratio of the sensor: So even if an extra two bits gives you an extra two in the shadows for the same quantization error, its difficult to believe that Canon have improved the signal to noise ratio by two stops [6 Db].
So: In practice you might be able to deliberately underexpose by a stop or even two to give youself over-exposure latitude with less of a penalty in the shadows - BUT only if you expose raw images and commit yourself to correcting for the underexposure before conversion or produce 16-bit per channel output.
For a correctly exposed image you are still going to be most concerned about the top six stops and eventually convert the image to 8bits oper channel for printing or display. A 12 bit raw image already provides more headroom than you would ever need for heavy processing [save for some Microphotographic or astronomical applications].
The practical cvalue of 14 bit raw files is very limited.
...Show more →
Do we know that earlier cameras (ie the 5D) were limited by the 12 bit ADC?
I do some HDR shooting, so 12 bits is helpful, you just have to increase local contrast.
The fundamental limit is the signal to noise ratio of the sensor: So even if an extra two bits gives you an extra two in the shadows for the same quantization error, its difficult to believe that Canon have improved the signal to noise ratio by two stops [6 Db].
So: In practice you might be able to deliberately underexpose by a stop or even two to give youself over-exposure latitude with less of a penalty in the shadows - BUT only if you expose raw images and commit yourself to correcting for the underexposure before conversion or produce 16-bit per channel output.
For a correctly exposed image you are still going to be most concerned about the top six stops and eventually convert the image to 8bits oper channel for printing or display. A 12 bit raw image already provides more headroom than you would ever need for heavy processing [save for some Microphotographic or astronomical applications].
The practical cvalue of 14 bit raw files is very limited.
...Show more →
Do we know that earlier cameras (ie the 5D) were limited by the 12 bit ADC?
I do some HDR shooting, so 12 bits is helpful, you just have to increase local contrast.
Pondria wrote:
There seems to be an expectation about finer tonal levels than 12bit. It will Not be the case automatically. What matters is where the sensor noise level is. With 1DsII, it has 9 stops of DR. In other words noise, level is 3 bits. If you increase the data size to 14bit without improving the noise, you get the same 9 stops of DR with 5 bits of noise, which is basically the theoretically minimum tonal resolution possible.
So, the effectiveness 14bit will be realized only if the noise level has been improved. How is it compared to 1D2 in terms of the noise ?
I disagree with you here because even if there is zero additional dynamic range between maximum sensor level and the noise floor, there will always be two extra bits in each ev interval providing four times as many tonal levels within that range. There will also be four times as many tonal levels for the noise but that is irrelevant. The brightest ev will have 8192 levels instead of 2048, and the ninth ev will have 32 levels instead of 8. (These numbers refer to the sensor levels and not to an 8-bit jpeg file or a 16-bit tiff file.)
For those who don't understand where these numbers are coming from, a 14 bit sensor can have 16,384 tonal levels and a 12-bit sensor can have 4,096 levels (2 to the power 14 and 12 respectively). The sensors are linear devices and so halving the light halves the tonal level recorded for that light. Starting at maximum level (whatever amount of light that may be) each ev or stop darker halves the level. So if maximum is 16384 then one stop darker is 8192 (giving 16384-8192 = 8192 levels in the brightest stop), another stop darker is at 4096 (4096 levels in that second brightest stop), another stop darker is as 2048 (2048 levels in that third brightest stop), and so on until you get to just 1 level in the 14th darkest stop. At that stage it's all academic anyway as the noise is far greater than 1. In the camera the levels will range from 0 to 16383 instead of 1 to 16384 because computers start counting at 0 instead of 1, but the principle I have described is unaffected by this little detail.
DrPablo wrote:
Yeah, scanners are a good example. Scan in 48-bit mode and all you get is a bigger file. Drum scanners can achieve exceptional resolution and with a noise frequency that is far below the frequency of film grain, so they really are delivering more detail; but an Epson 4490 or something is not delivering much more to a 16-bit file than to an 8-bit file.
There's little point in a huge bit depth unless your sensor is actually delivering more information into it.
On a decent slide scanner such as my Nikon 4000ED there is certainly more than 256 tonal levels extractable from a decent slide. That makes a 48 bit file more useful than a 24 bit file. Sure, there's a lot of noise too but there is extra detail in there. If a scanner - or rather all scanners - were incapable of producing more than 256 total levels then so too would digital cameras, and that simply isn't true. You would have seen that by now if you ever shoot in raw mode.
The higher end scanners are definitely better than only 8 bits, but no one could find an observable difference between the 12 bit scanners and the newer 14 bit scanners. Yet the mfgs quote the Dmax for those scanners as the theoretical maximum for 14 bits, not a real measurable level. Does the 14-bit / "4.2 Dmax" Coolscan V scan deeper into transparency shadows than the 12-bit / "3.6 Dmax" Coolscan IV? Not that anyone has been able to demonstrate. Thus one might think the noise limit was higher than the 12th bit, and the boost to 14 bits was for marketing reasons. This is easy to test with scanners - you have your film on the light table, and can see detail that neither scanner can see, and directly compare your old scanner to your brand new one. Search the archives at photo.net, there were a lot of these tests run when the 14 bit scanners came out.
pretty clear parallel here. cheap flatbeds, that's a different story where 8 bits probably is just as good as 16.
"