Alan321 wrote:
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.
Correct me if I'm wrong here; but because it's linear and not asymptotic, there will be an absolute floor and an absolute ceiling to recordable light intensity. However powerful the camera's processor and whatever the bit depth, this isn't changing the absolute floor and ceiling of the recorded light. Thus, the dynamic range doesn't change, even if the recorded data becomes more malleable (or malleable with more impunity).
hankk wrote:
Sorry, didn't see the second page. This is in response to Alan321's post discussing the values of the pixels.
This is interesting. If the middle tone is 17% gray, then the middle value (1260) yeilds a 100% value of almost 7412. (1260 / 0.17 = 7411.7)
If you multiply 16,384 by 0.17, you get about 2785. (the theoretical middle tone?)
if you divide 1260 by 16,384 you get 7.69%
Am I doing something wrong here, or do I not understand something?
It's interesting for us technical types but it's trickier than you realise. For starters there is debate about whether neutral middle tone lighting is 18%, 17%, 14%, etc. Then there is what it means anyway - the amount of the incident light reflected off a middle tone subject. That is not a fixed amount of light. It is relative to the intensity of the source light. So maybe it applies at noon on a clear sunny day on a particular date in a particular location. I don't know, but the light meters all seem to know what it is
The middle tone recorded by the camera sensor assembly is defined by the manufacturer to be anything they feel like making it. Pick a number, any number. It's just a number and it need not be in the middle of the sensor light receiving range. They could make the middle level on a 1D3 the same as on a 1D2 at about 370 or whatever, but that would simply cause the meaning of maximum level (16383) to be 6 stops brighter than middle instead of 4 stops. Whether the light measuring part of the sensor can actually handle such light levels is another factor to be considered by the manufacturer - they don't want to waste the data on unattainable levels.
The significance is that having chosen a middle tone level they calibrate the sensor output so that middle tone produces that level. Then it is just a case of simple mathematics to figure out how bright the maximum recordable light level is, and how dark the minimum recordable level is. These mathematical limits may or may not correspond to featureless white and featureless black.
My observation in DPP is that Canon still allowed a margin of a bit less than 4 stops between middle level (as defined by them) and maximum level (as defined by what a 14-bit analog to digital converter can record = 16383). Thus all of the extra data bits are being used at all light levels rather than just being used for two additional stops at the bright end. The middle level is therefore now at about 1260 on the 1D3 raw data but that has nothing to do with multiplying the maximum possible level by 17% or 14% or 18% reflectivity that we associate with middle tone. Any similarity would be coincidental.
Consider this: Canon could just as easily have used a 16-bit analog to digital converter and still left the middle tone level at 1260. It just so happens that if they did that then unless the actual sensor could count photons above a light level that is 4 stops above middle, the output would never exceed the equivalent of 14-bit output. That would make the use of a 16 bit AD converter an unnecessary expense.
It is only after the 8, 10, 12 or 14 bit sensor data is converted to a 16 bit raw data value that adjustments can be made in software or firmware to utilise the extra data bits in the file format. The original sensor is still limited to whatever the light-gathering and measuring technology allows. Previously that was equivalent to 12 bits but now it has been raised towards 14 bits.
I'd love to see a 1D3 high ISO noise reduced raw file so that I could see what the raw image data histogram looks like in DPP but so far I have not come across any. I don't have a 1D3 to take my own photos but I may try to test one out at the shop this week with my own memory card. If I can do that then I'll report back later on.
Alan321 wrote:
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.
I disagree right here.
Your tonal resolution is limited by noise regardless of the EV level. You somehow give us the impression that the finer resolutions will kick in only in the low signal area. That is not the case. The minimum resolvable step size between the levels are 32 ( assuming your 5 LSB's are noise ) regardless of the EV level.
DrPablo wrote:
Correct me if I'm wrong here; but because it's linear and not asymptotic, there will be an absolute floor and an absolute ceiling to recordable light intensity. However powerful the camera's processor and whatever the bit depth, this isn't changing the absolute floor and ceiling of the recorded light. Thus, the dynamic range doesn't change, even if the recorded data becomes more malleable (or malleable with more impunity).
You are correct that there are absolute ceiling and floor levels inherent in the data bit limits. However, those limits probably (not necessarily but almost certainly) exceed the measuring capabilities of the sensor - the part that feeds data into the amplifiers that feed the analog to digital converter to produce a number representing the light level. There is plenty of scope for the range of sensor measuring capability to increase and still be within the confines of the data container.
The maximum level is easy to imagine, but the minimum level is a little harder. Obviously the lowest recordable level is zero but how much lower than middle tone is that ? The answer is determined by counting how many times you can dividing the middle tone value (and subsequent results) by two until the result can only be expressed as zero. Not one. Not any decimal value. Just zero. e.g. from say 1260 we will go down to 630 (-1ev), 315 (-2ev), 157 (-3ev), 78 (-4ev), 39 (-5ev), 20 (-6ev), 10 (-7ev), 5 (-8ev), 2 (-9ev), 1 (-10ev), 0 (-11ev). There can be no -12ev, -13ev, etc. It stops at -11 because Canon had arbitrarily made middle tone a value of 1260 (approx).
We mustn't confuse the dynamic range of the container data (14 bit = 14 stops, or 16 bit = 16 stops) with the practical dynamic range of the sensor package (probably 9 stops or more by now, above the noise floor).
Have you noticed that on DPReview they do tests of dynamic range on jpeg files and usually come up with 8 stops ? Of course, because it's an 8 bit file there can never be more than 8 stops of DR. Mathematically impossible. ANything more than 8 is an indication of accumulated measurement or calculation error. Every decent DSLR sensor can do better than 8 stops, so they should only be testing converted raw data in 16 bit TIFF files to remove the file format limitations. No matter how much genuine DR the camera produces it will be limited to 8 stops once we convert the data to an 8 bit jpeg image file.
When we use 16 bit data files we have a potential DR of 16 stops and indeed we can achieve that very easily - just set one pixel to 0 and another to 65535 and you have a 16-stop dynamic range file. Easy! We can make it noise free too by changing everything below middle tone to pure black. Of course the image will look like crap and there will be little or no resemblence to what the sensor captured but the specs are impressive. This is an extreme example of how we can make any camera produce noise free images with great dynamic range at any ISO - just butcher the data and sacrifice the image details. That's what happens with most non-1-series cameras.
The 1D3 sensor and firmware can apparently produce a large dynamic range and low noise without sacrificing image detail. That's why it is so impressive. Bird photographer Arthur Morris has written that in the high-ISO NR mode the 1D3 3200 ISO images are as clean as 1D2 ISO 400 images ! I have no doubt that he would not be favourably impressed if the fine details were being discarded to achieve this, especially as he uses high resolution lenses and 1Ds2 bodies.
Pondria wrote:
I disagree right here.
Your tonal resolution is limited by noise regardless of the EV level.
Yes.
You somehow give us the impression that the finer resolutions will kick in only in the low signal area. That is not the case.
That was the wrong impression and not what I thought I said. The finer resolutions will be in every zone (stop, ev, whatever you want to call it), but it will surely be most apparent in the darker zones where there is a great deficiency in available discrete levels in a 12-bit sensor output.
The minimum resolvable step size between the levels are 32 ( assuming your 5 LSB's are noise ) regardless of the EV level.
I don't understand this statement. 32 will be the level below which there is predominantly noise and above which there is clean signal. However, each subsequent level is simply the next integer value up the scale until you reach the maximum. The step size is 1 rather than 32.
Let's assume that the 5 least significant bits of the 1D3 data are all noise, as are the least significant 3 bits of the 1D2 data. That leaves 9 bits to effectively describe the useful data for both cameras. But for the 1D3 they are the 9 most significant bits of a 14 bit value instead of what would be the 9 most significant bits of a 12 bit number for the 1D2.
Yes the minimum clean raw signal level may be 32, but one stop above that is 64, so there are 32 levels within that darkest stop of clean signal data. On the 1D2 the equivalent noise floor was at a level of 8 and there were only 8 levels available to describe anything in the darkest stop of clean signal data. So the 1D3 has 4 times as many discreet levels in any of the corresponding light zones (stops).
This is true in every zone, from darkest to lightest. The fact that there are 8192 levels instead of 2048 in the brightest zone is not nearly as significant as there being 32 instead of 8 in the darkest zone because 2048 is already heaps whereas 8 is not nearly enough. It's those extra 24 levels in the darkest zone, and 48 in the next zone, and so on, that will most reduce the visibility of posterisation and so give us smooth tonal gradations.
If you increased the exposure of the darkest clean stop in post processing then there would still be only 8 distinct levels for the 1D2 (even though they are now brighter than they were) and 32 for the 1D3. At this brighter level the posterisation of the 1D2 would be far more evident than when the data is all darker, and the 1D3 will look much smoother than the 1D2.
Its a good thing normal people don't read threads like this because they would think ya have to be an electrical engineer or nuclear physicist to be able to take a decent shot with a digital camera.
Now please excuse me since I have to head over to Walmart to buy their remaining inventory of throwaway film cameras before some other poor b-tard who stumbles across this thread buys them all
Alan321 wrote:
Have you noticed that on DPReview they do tests of dynamic range on jpeg files and usually come up with 8 stops ? Of course, because it's an 8 bit file there can never be more than 8 stops of DR. Mathematically impossible. ANything more than 8 is an indication of accumulated measurement or calculation error. Every decent DSLR sensor can do better than 8 stops, so they should only be testing converted raw data in 16 bit TIFF files to remove the file format limitations. No matter how much genuine DR the camera produces it will be limited to 8 stops once we convert the data to an 8 bit jpeg image file....Show more →
I'm clearly missing something here.
What does a stop, which is a relative statement about light intensity, have to do with bit depth, which is an absolute statement about digital encoding?
These two concepts are many steps removed from one another.
You can record 10 stops of light with a 4 bit device, and you can record 2 stops of light with a 16 bit device. In the former case it will be posterized if that 10 stops is a continuous gradation (because the transition from black to white is always stepwise in digital devices). In the latter case you will divide the total logE over the 16 bits of data, meaning it will either be a low contrast image with very fine intermediate tones (like the original scene), or it will be stretched into a high contrast scene by assigning a blackpoint and whitepoint to the limits of the scene brightness range.
But if you take a sensor that can respond to a 15 stop range simultaneously (rare though that situation is), then there is no reason why it can't perform an A : D conversion that encodes it in 8 bits.
A stop is not a digital concept. It's barely even an analog concept. It's purely a statement of relativity. Bit depth is completely different -- it has no analog corrolary, it's an absolute statement about data content, and the nature of that data content depends on the sensor's conversion and not on the ambient light.
John Power wrote:
Its a good thing normal people don't read threads like this because they would think ya have to be an electrical engineer or nuclear physicist to be able to take a decent shot with a digital camera.
Hang in here. I'll bring in Lawyer's cut very soon
I promise you.
Personaly I will take whatever the sensor is capable or giving me. If it is noise in the shadows, so be it. Let ME fix the noise, just give me a RAW scan of the sensor with no camera adjustments. Just scan the sensor with a 20 bit
ADC and dump the scan to a file and as the sensors get better and better the output will also. There are software and DSP engineers that are a hell of a lot smarter than Canon's software engineers that can deal with the noise. I'm not one of them, but I have seen them at work. Most photographers don't have a clue as to what is available in the scentific world. Of course part of the problem is that Canon is dribbleing the improvements that are available in the name of "new model" every 18 months.
What does a stop, which is a relative statement about light intensity, have to do with bit depth, which is an absolute statement about digital encoding?
In this context the term "stop" is the relative range of light levels that are 1ev apart - i.e. a range in which the brightest level is twice that of the lowest level, but in which there is no restriction to what the levels are. You could also call it a zone or interval or whatever. The number of such zones that you can measure with a linear device is directly limited by the bit depth of that device.
You can record 10 stops of light with a 4 bit device,
no you can't, if it is a linear counting device such as an image sensor that counts in incremets of 1. Remember that I am referring to the linear sensor data values where the light is captured and encoded into raw data - not to an external image file format where non-linear data mapping had occurred to make the data range fit. That's where the 12-bit vs 14-bit change has occurred in the 1D3.
On a 4-bit device there only 16 discrete levels available and the top one is 16 times brighter than the lowest one. A 10 stop range does not mean that the brightest value is 10 times the darkest value - it is 2^10 or 1024 times brighter. You cannot fit such a big value into a 4-bit number.
John Power wrote:
Its a good thing normal people don't read threads like this because they would think ya have to be an electrical engineer or nuclear physicist to be able to take a decent shot with a digital camera.
You could be right, John, and I apologise to anyone who thinks I'm trying to blind them with science. I was trying to explain why they can expect to get a tangible benefit from using the new camera instead of the older models, rather than simply tell them that there is a benefit. If it's beyond their grasp at the moment then that's not a problem - they must simply rely on the old "show me" approach, but so far there are still not too many examples out there to show them. That's why this thread was started
How about a quick diversion and then you guys can go back to your technical discussion. This will save John from needing to go to Walmart and it will prevent my head from exploding.
Does having 14bit files produce better real world images? By this I mean, will someone looking at a print 11 x 14 or larger, made on a good printer actually see a difference?
1. the system is limited by the weakest link. this may be the sensor or the ADC.
2. A 14 bit ADC is capable of defining a range that is four times the range that a 12 bit ADC can define. (and 2 stops is 4 times the linear values)
3. There is some noise in each pixel. Some of this noise is inducted from nearby traces on the chip, and some of it is due to the random nature in which photons hit the photosite -- for a rough approximation, think of rain on your yard--there may be 1 square millimeter that never gets a drop, and 1 square millimeter that gets 10 drops. But each square meter gets about the same amout of water.
4. If the noise is less than what the least significant bit (LSB) defines then the system may be limited by bit depth. If the noise is greater than what the LSB defines, the system may be limited by sensor noise.
5. JPG and other compression may also limit the system, but we can ignore them for this discussion.
6. I assume that the 1DMIII has less sensor noise than the 1DMIII. The 5D has less sensor noise than the IDMII but more sensor noise than the 1DMIII.
7. I assume that the 1DMII is limited by sensor noise, but the 5D may be limited by bit depth--that is, a 14bit ADC won't make a difference on the 1DMII, but may show a slight improvement on the 1DMIII.
8. The 1DMIII may be limited by sensor noise, but the limit is 13 bits (or somewhere between 12 and 14 bits) -- if it had a 12 bit ADC, the bit depth would be the limit. Canon may have gone to 14 bits because they wanted a little room in DIGICIII for beter sensors in future cameras, becasue odd numbers of bits are hard to fabricate, or becasue 13 bits is unlucky :-)
Hrow wrote:
How about a quick diversion and then you guys can go back to your technical discussion. This will save John from needing to go to Walmart and it will prevent my head from exploding.
Does having 14bit files produce better real world images? By this I mean, will someone looking at a print 11 x 14 or larger, made on a good printer actually see a difference?
Your print is battling to show a 5 ev dynamic range, so a well exposed image without much contrast from any camera will look just about the same in terms of tonal gradation (resolution is another matter). However, a not so well exposed image that was too dark and that required more adjustment in levels or curves to reveal shadow detail can and probably will look noticeably better in the print.
The visible differences that are due to the extra two bits of data capture are all in the dark areas - you won't notice the changes that are in the bright areas.
hankk wrote:
What I gather from this discussion is:
...
I think hankk captured it pretty well - even John P should be able to understand I would like to add that the current generation top cameras ( 5D, 1D2, 1Ds2 ) are utilizing about 9 bits above the noise. I just can't believe that 1D3 will bring in 2 or 3 stop noise improvements. 12 bit should be adequate. 14bit may be a ground work for their future road-map.
Alan and Dr. Pablo need to sort out the Stop vs Bit issues between them.
"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."
Short answer: No [mainly because of point 6 below]
Long Answer
1] In one sense a 12 bit histogram gives 12 atops of dynamic range. Not all that can be used because two constraints operate in the shadows. Sensor noise and quantization error. The lower useful limit depends on what quantization error makes a "just noticeable difference" and whether this is above or below the noise floor.
2] In most practical images you are mainly interested in the six stops of dynamic range below the highlight because of the limitations of the printer or output device.
3] You might be interested in digging deeper into the shadows if you want to recover shadow detail from images with very dark features or under exposed
images. In these cases an increased colour depth will only make any difference if the shadows are limited by quantization rather than other sources of noise.
4] If you have "exposed to the right" bit depth will make no difference to over exposure latitude: Because you have exposed to the right the over-exposure headroom will be exactly the same. You might get away with deliberate underexposure to give yourself some overexposure latitude.
5] In the six stops below the highlights, working with 12 bits per channel gives you 64 levels in the sixth stop - which is more than sufficient to protect you from "combing" or other quantization artefacts when processing images.
6] In fully exposed, 12-bit per channel image the top 4 stops have 2056, 1024, 512 and 256 levels respectively. This allows sampling which is already much smaller than the manufacturing variation from pixel to pixel [the evidence is the noise you can see even in correctly exposed highlights] so reducing the sample size by increasing the colour depth produces no more useful information.
To sum Up:
As far as I can see increasing the colour depth would only be useful a] If it was accompanied by a genuine increase in the sensor's signal to noise ratio and b] you make a habit of underexposing to give yourself some over-exposure latitude or c] to recover some shadow data from underexposed images or those with an exceptionally high dynamic range.
As for HDR I can't see that it would make any practical difference if the HDR images are generated by combining a set of standard images with exposures roughly two stops apart [there is not much point in using smaller gaps]. 12 bits per channel are more than enough for this purpose [8 bit images are definitely not - you get posterisation if you try and tone map them]. If you are thinking of one of the applications which tries to use tone mapping to recover highlight or shadow detail from single raw files of subjects with a high dynamic range extra colour depth might help if the signal to noise ratio has commensurately improved.
Alan321 wrote:
In this context the term "stop" is the relative range of light levels that are 1ev apart - i.e. a range in which the brightest level is twice that of the lowest level, but in which there is no restriction to what the levels are. You could also call it a zone or interval or whatever. The number of such zones that you can measure with a linear device is directly limited by the bit depth of that device.
Again, I don't see how this makes sense in an absolute kind of way. What matters is the brightness range over which the sensor is responsive. An infinite bit depth is useless if a sensor is only responsive to light over an 8 stop brightness range.
If you have a 2-bit device, the only values you can have are 00, 01, 10, and 11. If you expose it to a 10 stop scene, and have an appropriate A : D conversion algorhithm, you can place anything below zone 2.5 at 00, zone 2.6-5 at 01, zone 5.1-7.5 at 10, and anything above zone 7.6 at 11. So you've encompassed the entire dynamic range from black to white -- with a linear conversion. Of course you'll have extreme posterization.
But you can conceive of the same thing with an 8-bit device and a 12 stop scene. Assuming your sensor is actually responsive to a 12 stop SBR, then you can linearly convert your scene brightness in a way that corresponds to 1.5 stops per bit.
What doesn't make sense to me in this conversation is that the whole concept of a stop is really only useful in terms of optical parameters (and maybe ISO). The division of light into doublings, as opposed to any other kind of division, is done mainly for convenience and not because anything truly meaningful is meant by that ratio. Why talk about doublings of light? Because it's easy to engineer aperture and shutter speed steps that accomodate doublings.
But there is a continuum of light intensity from pure black all the way to the brightness of the sun, and the imposition of a 'stop' on it is completely arbitrary -- it says nothing about where on that continuum of light intensity you are.
On a 4-bit device there only 16 discrete levels available and the top one is 16 times brighter than the lowest one. A 10 stop range does not mean that the brightest value is 10 times the darkest value - it is 2^10 or 1024 times brighter. You cannot fit such a big value into a 4-bit number.
I realize that, but again, in the real world there are no discrete levels of light! It's just light, it's analog. The 1024-fold boundaries of this range need not be accomodated by 16.8 million values, or 256 trillion values, or 65,536 values.
Because light is continuous, not discrete, you're always splitting it into values on a digital device. So the tiny subdivisions of any given light intensity may be assigned to a higher or lower pixel value depending on where they fall. In other words, you're always creating a finite number out of an infinitely divisible continuum.
So the bit depth only determines the number of intermediate values where you can assign finer gradations of light. If you have a wider dynamic range, say 12 stops, what's wrong with 1.5 stops per bit on an 8-bit sensor?
In other words, you lose fine intermediate gradations by stretching the blackpoint and whitepoint farther apart; and with a low DR scene you add intermediate gradations by pushing them closer together.
All of this is consistent with a linear conversion. It has only to do with the slope and the intercepts of the line.
Alan321 wrote:
Your print is battling to show a 5 ev dynamic range, so a well exposed image without much contrast from any camera will look just about the same in terms of tonal gradation (resolution is another matter). However, a not so well exposed image that was too dark and that required more adjustment in levels or curves to reveal shadow detail can and probably will look noticeably better in the print.
The visible differences that are due to the extra two bits of data capture are all in the dark areas - you won't notice the changes that are in the bright areas.
As far as I can see increasing the colour depth would only be useful a] If it was accompanied by a genuine increase in the sensor's signal to noise ratio and b] you make a habit of underexposing to give yourself some over-exposure latitude or c] to recover some shadow data from underexposed images or those with an exceptionally high dynamic range.
Another good summary. Thank you. And we keep coming to the same conclusion -sensor S/N improvement needs to be accompanied.
Greenfield made another great point - In real shooting, by under exposing the shot, you will immediately lose your DR. Think about this. 1/2 of your data are in the top one stop. If you under expose by 2 stops, you are left with 1/4 of the data.
greenfields wrote:
2] In most practical images you are mainly interested in the six stops of dynamic range below the highlight because of the limitations of the printer or output device.
I disagree here. I've heard that printers give you about a contrast ration of 450:1 and monitors typically are with the range of 350:1 to 1000:1 256:1 is 8 bits, so we're talking about 8 or 9 bits (and stops)
Also, local contrast enhancement changes things--the levels of the output picture may be different based on the luminance of the surrounding areas.
If you look into this further, many of the "old masters" like Rembrandt and Vermeer used a sort of local contrast enhancement because their paints were as limited as our printer inks are in terms of dynamic range.
So you can only print x stops, but you may have one region on the input image that is brightened more than another region -- think about an HDR of a room-one shot for the room, one shot for the view through the window. You can get one image that shows both the outside and the inside nicely (this trivial case can be done in PS with a levels adjustment and cut and paste.)
The more bits of input, the better. But you have to process them correctly to get a good looking picture. (and there's a big difference between one of my feeble attempts and one of Rembrandt's pictures )
