There are two samples from the mkIII on the Imaging Resource website. They contain the black-pixel borders if you load them into the current version of IRIS (www.astrosurf.com/buil). Future versions of the program will most likely crop them away.
But if you have the extra 1-2 stops DR in the shadows can't you "underexpose" by 1-2 stops without sacrificing shadow noise? Thus you will gain 1-2 stops in the highlights without getting any worse shadow noise.
Yes. Dynamic range at capture (i.e. linear values digitized into a raw file) is agnostic of the intended exposure. Mid-gray has no inherent meaning in terms of captured and counted electrons. The downstream processing subsequently produces an output file in a generic colorspace, where mid-gray has a meaning.
That helps, but if there is a lot of read noise, photon shot noise is the least of your problems in the deep shadows. Read noise is the dominant noise at some point in the shadows, and below. The point where read noise dominates is higher in sensor wells with high capacity.
--This gives you grief with the highlights blowing out, hence the
need for a large full well depth.
The raw capture is digitized as a linear count of the charge
accumulated in each photosite. This is proportional to the number
of electrons stored. The number of electrons stored is
proportional to the number of incident (and captured) photons.
Because the raw capture is a linear count directly related to
incident photons, the dynamic range of the captured image can be
limited by having too few bits available to digitize the available
signal information. 14 bits stores two more stops of sensor data
than 12 bits. Fact. Not fiction. That's just how it is.
In the least significant bits, quantization errors add noise. With
bigger signals the digitization noise is insignificant compared to
the inherent photon noise in the signal. At each signal level,
there is a determinable photon shot signal to noise ratio. There
is a measurable camera/digitization noise floor. There is a
cut-off point at low signal levels after which the camera's readout
noise swamps the signal.
The sensors in the last generation of big pixel Canon dSLRs (1DmkII
and 5D) have large full well capacity at ~ 80000 electrons. If
you are utilising the full well depth (ISO 100 typically for 5D and
a bit less for the 1DmkII ref. PIXSurgeon), then the least
significant bit of a 12bit RAW file reflects the charge of20
electrons. This lowest level of information is subject to dither
(intentionally added digital noise) to avoid artifacts. The photon
shot noise at 20 electrons signal is equivalent to a bit less than
5 electrons. In other words, the digitization itself is
introducing worse noise than the noise inherent in the original
signal. Having 1 or 2 more bits available to the D/A conversion
would add useful dynamic range in the capture.
I assert that 12 bit D/A holds back the 5D and the 1DmkII. I have
an experiment that shows this.
This sensor-native DR capability on previous generation cameras is
made available by use of the ISO setting, which alters the analogue
amplification of the output signal. I have taken an "HDR" set of
images with altered ISO setting (100,400,1600), but constant Tv and
Av. I have an example HDR image here:
http://www.seeminglyabsent.co.uk/2007_03_18_hdr/htmls/0000.html
(Gallery of all shots is here:
http://www.seeminglyabsent.co.uk/2007_03_18/htmls/IMG_8766.html
I have equalized the three contributing shots for brightness in
post-processing, but look at the shadow noise in the dark window
reflection and the highlights on the tiles on the right (gallery
for shadows is here:
http://www.seeminglyabsent.co.uk/2007_03_18_shadows/htmls/0000.html
; gallery for highlights is here:
http://www.seeminglyabsent.co.uk/2007_03_18_highlights/htmls/0003.html )
Crops for shadows (ISO 100 -> 400 -> 1600)
Crops for highlights (ISO 100 -> 400 -> 1600)
By my reckoning, the improvement in shadows from ISO 100 to 400 is
dramatic. This two stop variance in ISO setting would be available
to a single shot with 14 bit conversion. The sensor has had this
capability for a generation already; extra bits at the D/A stage
are now making it possible to access this capability in
post-processing of a single shot.
So. I would assert that 14 bit raw has the potential to add DR
capability. If the full well depth of the 1DmkIII is still of the
same order of magnitude as the 5D and 1DmkII, then the 14 bits will
result in greater DR. The need to swing ISO to access this DR will
reduce. The exact DR will depend critically on exposure.
That's what I concluded a couple of years ago after doing similar experiments, however, I now have a completely different conclusion.
That does not really follow; your experiment (and the ones I did with my 20D) show clearly that there are small-signal details in the sensor wells that are not captured well in the RAW files at ISO 100. However, more bit depth will not really help, because it isn't the problem. The problem is analog read noise. Most of the read noise in the ISO 100 shadows occurs somewhere between the on-pixel amplification, and the digitization in the ADC. The characteristic of the noise floor is not quantization noise; it is horizontal and vertical banding, mixed with 2-D random noise. The standard deviation of ISO 100 blackframes from current Canons range between 1.26 ADUs and 2.07 ADUs. Even with trace amounts of analog noise, quantization noise will not give a standard deviation of more than about 0.2 or 0.3. There are ways to prove that the noise is not mainly quantization; when I did my tests like yours, I did everything in the RAW state, without any converters per se, and had the opportunity to do simple arithmetic on the RAW data. I took two identical photos, one at ISO 100, and one at ISO 1600, both under-exposed so that they were actually pushes to ISO 10,000. Both had the same manual exposure, same Av and Tv values, but the ISOs differed. In the ISO 100, 20 RAW levels represented the full DR in the push, and in the ISO 1600, 320 levels did (a factor of 16, as you would expect). The ISO 100, as you would guess, is much noisier. To determine if it was quantization, I quantized the ISO 1600 to only 20 levels also by dividing the RAW values by 16, and then using 20 RAW values also for the full push. The result was shocking (at the time; logical to me now); the quantized ISO 1600 looked almost indistinguishable from the non-quantized 1600, and still far better than the ISO 100. Conclusion: bit depth is not a/the problem; analog read noise is.
Unfortunately, the noise and DR of a pixel in a mkIII is exactly the same as the mkII, at least at ISO 100. The 2 extra bits made the read noise quadruple in ADUs, and not much else. For all intents and purposes, those extra two bits are just a waste of CF storage. Unless, of course, with these extra two bits comes a reduction of banding noise (especially at high ISOs). I haven't had a chance to examine high-ISO RAWs from the mkIII, only ISO 100.
There really isn't that much room for improvement in "shot noise". It's not a contamination of the actual scene; shot noise is what is really there in the real world light; we only call it noise because we don't like to see it. In real life, shot noise is buffered from the user by the brain. When photons are rarely striking individual rods and cones on your retina, your brain maintains a semi-fabricated "noise-reduced" image that is only changed locally as new samples are integrated. In a photograph; our brains take shot noise literally, as part of the texture of the medium, not as a quality of light, so it is not buffered from our perception.
That's only a semantic, relativistic issue, though. The bottom line is that the RAW data at any given ISO clips at a certain level, and a signal with what you consider the "lowest usable S/N ratio" lies a certain number of stops below that point; that is all that determines dynamic range. It is irrelevant and arbitrary where the middle-grey point is; it is only a frame of reference for metering.
5 is in the range of the best cases. Canon DSLRs generally have only about 3 to 6 electrons of read noise at the highest ISO, and A small ZLR like my FZ50 has a read noise of 2.7 electrons at ISO 100 (similar photon count, btw). At ISO 100, and 50, however, read noises on Canon DSLRs can be as high as 36 electrons. My FZ50 rises to 3.34 electrons at ISO 1600, in the opposite direction than the Canon DSLRs; the Canons are optimal at high ISOs, and the FZ50 is optimal at ISO 100 (because the ISO-gain amplifier is basically garbage in the FZ50; pushing gives better results).
But I'm up on what the meaning of "is" is,...and I'm right on top of those "and"'s and "the"'s.
I know it's always easy to be most influenced by whomever has spoken last, ...but my bottom line of the moment "take" on the Peter/Pix/John input is that the 1DIII has no appreciable "new" DR advantage.
Can we just be clear that the 36 electrons equivalent readout noise at ISO 100 (say) for a current Canon big pixel 12 bit capture is dominated by quantization noise.
Expose to the ... left, with greater sensor + readout DR allowing the latitude.
I think the greatest insight will come from real photographers in real shooting situations, learning the new capabilities of new equipment.
--
...while other components of readout noise will not change magnitude and will be seen as more dominating components in the overall lowered noise level.
