Can someone illuminate me on how the newer cameras can capture 14 Bit color but according to Canon's white paper the 1DS Mark 3 does NOT have any more dynamic range than the Mark 2. I am not enough of an expert to reconcile this apparent contradiction. Thanks

From a guy over at DPreview.com ...



Found here (you may have to cut and paste. Dpreview can be pretty fascist about direct linking.) ...

http://forums.dpreview.com/forums/read.asp...essage=24492312

An analogy: the dynamic range is the height from the darkest part of the valley below to the brightest light at the top of the mountain . The A/D converter makes the steps that lead up the mountain (14bit converter makes more steps than a 12bit converter) - more steps going up does not change the mountain, just makes for a smoother climb!

That point is often made, at one time by by no less than Bruce Fraser, but I do not think that it is always true. The matter is discussed by Roger Clark in his digital sensor performance essay. The increase from 12 to 14 bits can potentially add 2 stops of DR to the image.

This link shows why ADC noise is affected by bit depth. The difference would be most prominent at low ISOs with large pixel sensors.

This Another Link on DPreview in which I participated gives additional information.

Bill

The analogy was to illustrate the general point about the terms. However, I'd agree that if your system is limited by the ADC, it could yield higher DR when given a higher bit depth, but you need to start specifying more details about S/N and amplifier and read noise etc, e.g. this (from Roger Clark at your link - my emphasis).

" If 14-bit or higher analog-to-digital converters were used, with correspondingly lower noise amplifiers, the dynamic range could increase by about 2 stops on the larger pixel cameras. The smallest pixel cameras do not collect enough photons to benefit from higher bit converters. "

IOW, no question that there are more variables than simply bit-depth and DR in a complete discussion!

I accept that DR and bit depth are 2 different things.

A thought experiment - 10 stop DR and 2 bit depth. Wonder what and image under that scenario would look like?

So back to the ladder analogy. 10 stops - so say arbitrarily say 20 feet high. 4 bits, so 16 steps. Each step is 1.25 feet. So I have to be 1.25 feet off the ground before I can "see" into the next step. Likewise I have no was of seeing any difference after 18.75 feet.

Now take the same dr (20 feet) but at 6 bits, so 64 steps, each step is now .313 of a foot.

Now at the bottom of the ladder I can discern a difference between .312 and .314 of a foot, and similarly at the top I can get to 19.688 before I no longer able to discern a difference as I move up the ladder. So given 12 vs 14 bits, with 14 bits I'm able to extract a lot more information from the top and bottom stops, which "looks" like more dr.

Based on my short term experience with the 1d3, it feels like I'm getting more dr - in that I can dig more detail our of the shadows and highlights than I could with a 1d2 - but it's really the difference is bit resolving power.

Does this make any sense?

Edit: I also recognize that the distribution of "steps" is exponential, not linear, but the point of increased ability to quantize at the ends adds more usable range.

The marketing department has decided to fill your CF cards and hard disks with two bits of noise! Nice to know they're thinking about the users!

The 1Dmk3 definitely has more pixel-level DR at ISOs higher than 100; up to 1/2 stop more at ISO 1600. At ISO 100, however, it has almost exactly the same DR at the pixel level as the mk2. When you consider the entire image, however, with more pixels, the limits due to each pixel becomes lessened, so viewed at the same size and properly resampled (if necessary), the mk3 should be slightly better, even at ISO 100. Also, RAW converters may use more bit depth for conversion, reducing posterization of the output (but all these benefits could be had if the 2 extra bits were "10" for all pixels).

Isn't the key to understanding this connected with the concept of the linear progression? Each additional f/stop of dynamic range requires double the number of photons to impinge upon the photoreceptor.

There will be a minimum number of photons required to produce a recognisable image that is not completely or almost completely obscured by noise. Having decided what that number is, you can then keep doubling the number till the photoreceptor well is full. eg. (50, 100, 200, 400, 800....128,000, 256,000, 512,000 etc).

Once the well is full, no futher increase in the DR of the captured image is possible. Clearly, sensors with bigger pixels have a DR advantage, all else being equal.

It seems to me the technological challenge here is to devise a way of creating a non-linear progression which more closely matches human visual acuity. For example, the brightest stop of DR in the above progression (256,000 to 512,000) contains far more levels than the eye can discern (what a waste), whereas the darkest stop of DR (50-100) contains far too few levels for good tonality.

One might expect that 14 bit conversion would help improve tonality in the darkest parts of the image so that what was previously unacceptable with 12 bit processing now becomes almost acceptable. One thus gets a marginal increase in DR through slightly cleaner shadows, but to get a substantial increase in DR (say one full stop) one would need photoreceptors that can accommodate at least double the number of photons, say 1,024,000 in the above example. Ie., to get one additional stop of DR, each photoreceptor needs to have double the capacity, all else being equal. (It's always necessary to add all else being equal ).

Tim,

I'm glad you added that edit, because the linear encoding currently used by all raw camera files does place a strict limit on DR. With 12 bit encoding the largest encoded value is 4095 and the smallest is 1, giving a DR of 4095:1. Log base 2 of 4095 is 12, so the maximum DR is 12 stops. With 14 bits the maximum DR is increased by 2 stops. Of course, adding bit depth will not help DR unless all elements in the imaging chain are capable of supporting the increased DR.

It should be noted that even 16 bits is insufficient for HDR, and one must use log or floating point encoding, as discussed by Greg Ward.

Bill