"ISO3200 Special" Native CR2 + JPG test results
These were shot with the same parameters and conditions as the DNG+JPG in my previous post. To decode CR2s I used RawTherapee. Fig 1 is CR2 left column and parallel JPG right column. Shooting parameters are displayed in the thumbnails.
Figure 1
A note on barrel distortion first.
>> Before RT decodes a CR2 (ISO 400, no green checkmark), it displays the original thumbnail from the camera. Because then incoming thumbnail is barrel-corrected, this demonstrates that the EXIF thumbnail is a small version of the JPG, which we know now is a barrel-corrected CCD raw, done in the camera.
>> After decoding a CR2 (green checkmarks), RT updates the thumbnail. You can see the barrel distortion and more subject area captured in the CR2 raw image data when compared to its JPG.
>> Hence we can conclude that CR2 raw data as emitted by the camera, is not consistent with the thumbnail. So if one is using Canon PC software to decode and view CR2s, decoded CR2s get barrel-corrected on-the-fly on the PC, giving one the illusion that they are the same // nice trick. Hence this new data confirms that the PC software must know the camera model to apply the appropriate algorithm.
The foregoing raises the question: does the ISO3200 Special undergo a similar process? What really happens inside in the camera?
ISO3200 Special
>> In Figure 1, ISO brightness stepping from row-to-row is consistent. So we can firmly establish that after RT decodes the CR2, image exposure also remains unchanged.
>> Hence in that respect, ISO3200 CR2 brightness is as originally given by Canon, and is the same as the JPG, leading one to conclude there are no brightness correction-specific exposure transforms embedded in EXIF.
This begs the question then ... what is special about this high ISO mode? Does Canon expose at 1600 and contrast-push the exposure? Is this a true electronically-increased VGA gain setting? A combination? In Figure 2 seeks to find some answers.
Figure 2
Is Canon ISO3200 Special a fake? Here I expose the scene at ISO 3200 top left, and 1600, top right. Bottom is a color-balanced tonecurve-pushed version of the ISO 1600 I did by hand so that the wall background and keyboard areas have increased intensity by very close to 1EV, on the plus side // if you look carefully you see that the push is slightly more than 1EV, done intentionally to be on the safe side as you will see below. Hence if one part of the scene is exposed equally across the two images (wall and keyboard), all parts should follow (computer screen). That means if Canon indeed creates a 3200 by shooting at electronic 1600 and then does a trivial intensity change in firmware, the bottom image should be identical to the top left.
It is not identical ... observing the yellow zones, it is clear that the highlights are considerably higher intensity (around an additional 1EV) in the ISO3200 than the pushed ISO1600, despite the EV1+ on the 1600. I also find that the vertical wall transitions borders appear sharper in the ISO3200. And what of the noise and SNR? By observation noise appears roughly the same between the pushed ISO1600 and the ISO3200.
Some conclusions
That result leans towards a preliminary observational interpretation that the 1EV+ ISO 1600, noise seems to increase as expected, but in the ISO3200, the scene signal strength does appear to be higher, thus one could conclude the Canon "ISO3200 Special" is unique and not a fake. In my early empirical analysis of this camera I paid much attention to this setting. Results then were consistent with these, but with very dark scenes, full iris and very long exposures. I could even say that ISO3200 performs better at the longer exposures: overall, after taking many shots at 15 seconds it was clear to me that the 3200s had better SNR and image rendition than pushed-1600s under the same environmental and setting conditions. This means that with dark-subtracted long exposures (these here are 1/4s not dark-subtracted), ISO3200 Special performs very well. Perhaps Canon does change VGA amplifier gain for the 3200, but it is not exactly 2X over 1600, after which a sophistical noise-reduction algorithm handles the rest. It could be a combination of factors that give excellent results.