hankk wrote:
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.) ...Show more →
And this is true of master darkroom printers as well. Bruce Barnbaum is a good example these days. He shoots in extreme dynamic range settings, with as much as a 15 or 16 stop range, then uses severely contracted development (stand development with dilute developers) to pull as much as 7 or 8 stops. This works because highlight areas exhaust developer much more quickly than shadow areas, and with dilute developers (and minimal agitation) he allows the shadows to build density without the highlights getting out of control. So he's able to modify development to make a negative that's easily printable. Ansel Adams also used contracted development (when needed), but seldom more than 2 or 3 stops.
Dynamic range for film is measured using densitometry and plotting D/logE curves. The relationship between density and exposure is analagous to the digital relationship between signal and exposure. What Barnbaum is doing is essentially doubling the number of stops of scene brightness represented by the inherent density range of the negative. He's not changing the density range (aka the bit depth), he's only changing the slope of the conversion.
When I arrived I noticed that they were no longer stocking the disposable film cameras. I questioned the clerk and she recommended that I buy a camera called a Canon 1DMKIII.
The imperfection of optical surfaces, optical coatings, baffles and flocking are responsible for sending one photon per 1000 to the wrong place on the sensor. If displaced by a small amount, you see a halo or glow around a bright object. If displaced by a greater distance in an orderly way, you see flare. If displaced by a random process, you get veiling glare (a general reduction in dynamic range).
As a qualitative demonstration, observe all three effects in the following sequence. Lens used was the Zeiss Aposonnar 200/2 with T* multicoating and all the best reduction of stray light that $6000 can buy. Subject was a Maglite bulb in a dark office. By increasing exposure time, the existing low-order chaos becomes high-order chaos. Future sensors with higher dynanamic range will not reveal shadow detail, but will merely render the chaos with an ever-finer palette.
For those who want know the dynamic range of a given Canon sensor, simply dump out the raw bits with Dave Coffin's dcraw utility, and apply your favorite statistical measure to a selection of the pixels. At the limit, results will depend on the choice of test scene and its placement of light sources.
Pondria wrote:
So, your point is that 1:1000 or ~10 stop is the max DR regardless of noise or bit depth ?
Yes, although I mean to perform a quantitative experiment one of these days. It also means that HDR is a bright, shining lie. For modern optical multicoating, Schneider-Kreuznach gives a transmission rate of about 99.75% per surface in the green spectrum. That leaves up to 0.25% bouncing around in the optical path, and potentially landing on the sensor. The famous "UFOs" are caused by such a reflection between the sensor cover slip and a front filter (both planar). Dynamic range of the sensor is an issue of rapidly-diminishing importance - at least for Canon!
Pondria wrote:
Well, then, back to my very original question - what may be the benefit of 14 bits ?
From a single sensor site, Canon can surely achieve the DR of 1:16000 which justifies a 14-bit encoding done the dumb way (linear). Of course, they could encode the same data using fewer bits by any number of mildly-lossy compression methods (like a gamma curve). CRW itself implements a lossless compression. Frankly, I think "14 bits" is marketing since the MP race is over: bit-depth is the new sales frontier. Touché, Rudi!
Pondria wrote:
Well, then, back to my very original question - what may be the benefit of 14 bits ?
Processing headroom and prevention (further reduction) of quantisation errors in shadow parts.
If you put 10 stops into 10 bits, then the first stop is single bit (on/off) precision. Ouch. By adding 4 bits to the bit depth... well, you know the drill by now, it will have 16 tonal gradations.
Given the fact that some high-end monitors, videocards and software can do 10-bit output already, I'd say 10 stops (and 14-bit "room" to house those 10 stops) seems like a sweet spot.
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.
There is another flaw in the argument here apart from what others have mentioned about printing capabilities, and that is that Canon have already built in an almost 4-stop difference between middle tone and the maximum recordable highlight. So a 6-stop dynamic range would leave you only 2 stops below middle tone. That is not enough even without allowing for editing and processing - even slide film can do better
The simple answer is with 14 bit A/D, the noise floor is effectively lower than with 12 bit, therefore raising the dynamic range, which is defined as maximum signal divided by the noise floor. The Canon 1D Mark II and 5D are limited by 12 bit A/D; the 1D Mark III, even without its sensor improvements, would benefit from 14 bit A/D conversion.
Pondria, the benefit of 14-bit over 12-bit was not necessarily that there will be two extra stops of noise-free dynamic range, but that there will be more data values to describe the dynamic range that is captured. Less quantisation. Less posterisation. Potentially better prints after editing or manipulating shadow levels.
Dr Pablo, the issue at hand is not that any dynamic range could be mapped into any bit depth (of course it could be done by careful selection of a mapping algorithm), but rather that what the Canon sensors actually do is capture light, convert it to an analog voltage, amplify it and so on, convert it to digital and record it as a number in a linear range from 0 to the bit depth limit of the AD converter in increments of 1. Any other possibiities are valid but not applicable.
Adding extra bits to this technology allows for more data values to describe the useful light levels above the noise floor and that allows for potentially better images whether or not the real captured dynamic range is increased.
[for those who don't understand what I mean by linear, it is simply that whatever happens to the input also happens to the output - e.g. double the input and get double the output. Non-linear doesn't work this way - e.g. doubling the input may give only a slight increase in the output. The sensor is a linear capture device, but the subsequent conversion from raw image data values to 8-bit rgb values is non-linear. That's how a 12- or 14- stop range of raw data values gets mapped to just 8 bits in a jpeg file while still retaining the black, white and middle tones. Film is also non linear.]
What also comes into the mix is noise vs signal. It's not clean cut at all.
Qualitative analysis says there is a signal. This means that the signal itself has to exceed the background noise by at least double the noise itself.
So if you have a 1 volt deviation around 0, your background noise is between -0.5 and 0.5 volt. A signal has to be at least twice the deviation, more than 1 volt (or -1 volt) to be considered a signal. If not, it's within the noise.
Now, Quantitative analysis demands more. Of a signal of twice the noise value you can only say "it's there", you can't say anything conclusive on how high the signal is. Rule of thumb is that you need a signal of at least 4 times the background noise value (and even then your accuracy is plus or minus 25%).
This is also the reason why dynamic range testing is subject to a quality factor. If you use 4 times the statistic deviation on the signal as your lowest value, you get a higher DR at the cost of image quality. If you are more stringent and use 8 times the deviation as lowest usable value, then you come up with a stop less in DR. (But less chance of chroma artefacts, thus higher image quality.)
So this is another way of saying that a better signal-to-noise sensor needs more accurate read-out, otherwise intensity signals at the lowest end of the range remain meaningless and the increased SNR remains untapped.
"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.
There is another flaw in the argument here apart from what others have mentioned about printing capabilities, and that is that Canon have already built in an almost 4-stop difference between middle tone and the maximum recordable highlight. So a 6-stop dynamic range would leave you only 2 stops below middle tone. That is not enough even without allowing for editing and processing - even slide film can do better."
I think that's a good point. I was guilty of subliminally assuming that everyone exposes to the right just because I do. I can see why most users would rely on the meter for most of the time.
Why is it that many people can't separate the the issues. The sensor is an analog device that produces an analog voltage. Period! Thats all it does! Each cell produces a voltage between 0 volts and the sensor max, what ever that is.
Call it one volt for a talking point. When light hits the cell we now have a voltage that has some level of noise embedded in it, commonly called the noise floor. The quality of the sensor and the circutry set the noise floor for each cell. NOW that cell has done its job. the voltage is stored in a sample and hold curcuit and the cell is reset to zero. Its all done. forget the sensor from this point on. What the voltage represents is determined by how the sample and hold for each cell is processed. This is just a voltage, Thats all! it doesent have any color information or anything else in it,
What is known about the cell is it position under which color filter, so now the voltage is known to be produced by the amount of light through the filter.
Now the processing can start. With 8 bits you can divide the voltage by 255, with 12 bits 4095, with 16 bits 16,000+. With 16 bits or more, preferably 32 bits, you can now use DSP (digital signal processing) to extract signal 6 db below the noise floor. (Thats so low that all you see is noise on an ocilloscope.) Remember at this point its just a digital representation of an analog voltage. Thats all that is known about it! The accuracy of the voltage is how well it has been processed up to this point.
NOW, We scan this individule voltage produced by what ever the sensor size, Pick one. 18MP? We now have 18+ million separate voltages each representing a position under a color filter and let the cameras computer create a mosaic of the voltages.
Breaking it down you can see that the quality of the sensor is determined by how well the light hits each cell and how much splatter to ajacent cells. This is mechanical. Controled by microlens, filters, quality of the cell. etc.
The accuracy of the output file is controlled by how well the resulting voltage is processed. Damm that turned out long winded, even when simplified. But I hope it sheds some light.
That's been my point, Paul, (though less technically precise in my words).
The whole idea of a stop is long gone by the time you're dividing your data among different bits. And the absolute minimum and absolute maximum recordable brightness levels don't change based on how many divisions there are between them.
rico wrote:
According to an HP engineer, properties of the optical path dominate when you reach a dynamic range of 1:1000:
The imperfection of optical surfaces, optical coatings, baffles and flocking are responsible for sending one photon per 1000 to the wrong place on the sensor. If displaced by a small amount, you see a halo or glow around a bright object. If displaced by a greater distance in an orderly way, you see flare. If displaced by a random process, you get veiling glare (a general reduction in dynamic range).
As a qualitative demonstration, observe all three effects in the following sequence. Lens used was the Zeiss Aposonnar 200/2 with T* multicoating and all the best reduction of stray light that $6000 can buy. Subject was a Maglite bulb in a dark office. By increasing exposure time, the existing low-order chaos becomes high-order chaos. Future sensors with higher dynanamic range will not reveal shadow detail, but will merely render the chaos with an ever-finer palette.
[snip photos]
For those who want know the dynamic range of a given Canon sensor, simply dump out the raw bits with Dave Coffin's dcraw utility, and apply your favorite statistical measure to a selection of the pixels. At the limit, results will depend on the choice of test scene and its placement of light sources....Show more →
I notice that all the stray light is localized to one area around the bulb. As the light gets brighter, the region increases (so we may say that 1 in 1,000,000 photons travels a further distance...)
Dynamic range of greater than 1:1000 is important for those times when you have dark areas and light areas in your photo. A shot of the edge of a jungle with the sun high in the sky is one example. It may be quite dark under the trees, but the stray photons from the sky will mostly spill onto the leaf canopy, the leaf canopy will have fewer stray photons which spill onto the jungle floor.
Yes, local contrast may be limited to 1:1000 by the sensor, but different parts of the picture can have different local contrasts.
DrPablo wrote:
That's been my point, Paul, (though less technically precise in my words).
The whole idea of a stop is long gone by the time you're dividing your data among different bits. And the absolute minimum and absolute maximum recordable brightness levels don't change based on how many divisions there are between them.
IIRC, photons are binary-its either there or not. So you are measuring a number of photons with either film or sensor. The problem is, we can't count individual photons and so have to aggregate them. Film uses grain and density. Digital sensors use photosites and voltage levels.
There is noise in both methods, but film is not subject to EMI as the sensors are (and the sensors generate a lot of the EMI that causes noise).
Its probably better to say that film has many more discrete levels, and the processing can non-linearly amplify these levels to the range where the human eye can differentiate them.
On a sensor, there smallest difference is 1 (assume a scale of 0 to 16K), while in this same step ( 1/16K of the total range) film has thousands of steps (think of the difference between a 14 bit number and a 32 bit number -- both can have the same bottom and top level, but the 32 bit number has many more steps).
Let's say that you process both images (film and sensor) by multiplying by 16. Now the first 16 values of the digital image are the same, the next 16 values are the same and so on. But film will have different values for each step because there was a greater number of steps buried in the film. So the digital image will have some banding at the dark end. Noise will show up more becasue there will be a greater difference between adjacent pixels.
Paul Gardner wrote:
Why is it that many people can't separate the the issues. The sensor is an analog device that produces an analog voltage. Period!
...[how digital data sampling works]...
I think many people here understand how it works. You CAN represent the data with as fine quantization as you want. But the question is if it is effective. You can put the tick marks on the ruler at every 1/256 inches. But if the tick mark itself is 1/16" thick, what determines the resolution of the ruler ?
I agree that 14bit gives the processing headroom. But we are talking about RAW data, right ?
