If you are looking directly at the RAW data, and in fact there is additional data hidden in the RAW file how are you obtaining that data?
That's the whole point, you can't just look at the RAW data and see the data that Canon hid in it.
The RAW converter 'cooking' that you describe is Canon RAW converter accessing the hidden data.
You seem to think that Canon's RAW file format, CR2, is some kind of super top secret, encrypted data format that contains the deepest, darkest secrets that would somehow unlock the full potential of Canon's theoretical Nikon-comparable DR...
That is not the case. The CR2 format is a public specification (Canon even offers free source code that demonstrates how to read, write, and process their CR2 format), which is essential to supporting third-party development of RAW conversion and processing tools...such as Lightroom, Aperture, etc. Additionally...its a "RAW" file...it doesn't contain anything particularly special or super-secret that would need to be "hidden" that didn't come strait off the hardware...it has the full sensor readout...both the unmasked light-sensing pixels as well as masked "calibration" or "baselining" pixels around the margins, camera settings from the firmware, etc.
This isn't much different from any other manufacturers' RAW format, only its in a structure and form that allows Canon to achieve the things they need to...such as a high write out rate that supports the frame rates of their sports bodies. Canon, like many sensor manufacturers, also makes use of what is called a "bias offset" allowing a dynamic black level to be used when processing the RAW files. Nikon does not use a bias offset...their black level is fixed at zero, however with their low read noise, thats not an issue. The only real question regarding DXO's handling of CR2 files is whether they are properly factoring in the bias offset. Working in Adobe Lightroom, I wouldn't say that the 5D III is as bad off as it seems relative to the D800 based on DXO data...but there is
definitely a difference, and the D800 certainly has the edge.
Come on people, catch up, banding / pattern noise from every Canon camera, uniform 11 EV max from the unhidden RAW data across the board, and the CR2 file format can't have any hidden data?
Oh, and sensors that should have greater Dynamic Range like the 5D Mk III over the 5D Mk II have more read noise, and that read noise can't actually be data?
Read noise is an electrical byproduct in the sensor HARDWARE. A sensor is a signal receptor....it receives a two dimensional spatial waveform that must later be read out to produce a digital image. The process of reading that image signal introduces its own signal into the final product. There are a variety of forms of electronic noise that can interfere with an image signal, including dark current noise, sampling noise, differential noise, conversion noise, thermal noise, and noise generated by external fields interfering with the signal of the sensor itself. In the grand scheme of things, electronic noise, even in the 5D III which has very high read noise, is a small fraction of the maximum image signal (in the case of the 5D III, read noise at ISO 100 is 33.1 electrons, while maximum signal potential, or maximum saturation, is 67531 electrons...a ratio of 2040:1 at max. sat.!)
The problem with read noise is not when your image signal for a given pixel is high. If you have a nearly white pixel, read noise is 1/2040th the amount of the image signal...inconsequential. You are producing an output sample that is 99.9995% accurate. The problem with read noise occurs when the image signal for a given pixel is low. If you have a nearly black pixel, the read noise becomes a far more significant fraction of the image signal, possibly even overpowering the image signal. Again, with the 5D III, if the SNR of the shadows is say 1/2, then your signal is around 66e-, while read noise is 33e-. You are keeping a mere 50% of the image signal in the output sample the sensor originally registered...you've lost 50% accuracy. If your SNR in the shadows is 1/1, then your signal and your read noise are both 33e-. Now, you have nearly a 0% level of accuracy producing an output sample for that pixel. If your signal level is lower than the read noise level, it just becomes more useless noise.
Now, contrast the same electron levels above, only with a D800. The D800 has a roughly constant read noise level at all ISO settings of 3e-. Its maximum saturation is 44972 at ISO 100, which means its signal to noise ratio is 14990:1 in terms of electron charge. If the signal charge in a pixel is 66e-, instead of having a 50% image signal loss you now only have a 4.5% image signal loss. If the signal charge is 33e-, you have a signal loss of 9%. You also have the ability to sample an image signal that is considerably lower than what the 5D III is even capable of. At 10e-, your still only losing 30% of your image signal to noise. Only by the time you reach an image signal level of 6e- do you reach that 50% level that the 5D III reached at a signal level of 66e-!
Read noise...probably better termed "electronic noise" as its due to the electric and thermal properties of sensor during readout...becomes a part of the output signal from the sensor. (Note: Read noise isn't independent...its not something that can be represented separately in a RAW file...its an intrinsic part of the signal!) That output signal is then passed through an ADC, or analog-to-digital converter, which produces a digital file that, as closely as possible, represents "the signal" the ADC processed. Since electronic noise is a part of "the signal", it effectively gets "baked in" to the RAW file. Once its baked, there ain't no unbaking it. You can scrape off the charred parts, smooth over other unsightly lines and edges, and frost the top...but your cake is still gonna smell burnt if its burnt. ;P
