G9 studio scene comparison

[bobn2]

...

When you take a picture the noise in that image is not limited to only read noise.
So the PhotonsToPhotos approach is clearly more realistic than that of DxOMark.

I think you're over-egging the pudding when you say 'clearly'. Your PDR metric is not based on any perceptual research, simply your idea of what is reasonable. ...

I think it is rooted in perceptual research such as known characteristics of visual acuity.

Didn't you just apply a standard COC, in terms of the normalisation? In which case, it wasn't based on a whole load of research on acuity, just a few assumptions about print size and how long people's arms are. I'm not sure where the SNR of 20 comes from. See below.

It is true that within that framework I chose certain thresholds; also based on ISO definitions (like acceptable SNR)

ISO12232 doesn't mention and SNR of 20. They have 40 for 'excellent' and 10 for 'acceptable' for the noise based ISO speeds. The discussion in the standard is interesting:

The S/N values of 40 for the “first excellent” image and 10 for the “first acceptable” image were determined using subjective experiments performed during the development of this International Standard. These incremental signal-to-noise ratios were judged to provide “excellent” and “acceptable” quality prints of typical pictorial images using a high quality printer at approximately 70 sensor pixels/cm on the print (just small enough to be visually imperceptible) using normal tone reproduction. Note that 70 pixels/cm at a standard viewing distance of 25 cm corresponds to 30 pixels per degree of visual subtense. For prints made using significantly higher sensor pixels per centimetre values, lower S/N values may still yield acceptably low noise prints, while for prints made using significantly lower sensor pixels per centimetre values, higher S/N values may be required to provide acceptably low noise prints. In these cases, the S/N value for “excellent” quality prints is approximately equal to (70/P) times the S/N values listed, where P is the actual number of sensor pixels per centimetre on the print.

The last bit makes it all a bit vague, which maybe one reason why the noise based speeds never gained much currency. Is your '20' based on the scaling of the '10' for the CoC that you use? The 30 pixels per degree is based on optometry, the old Snellen eye chart (1862), and what is considered acceptable vision.

If that's the case, it's worth pointing out that the overall SNR of an image where the lower bound (the blacks) have a SNR of 20 is a lot greater. And just saying it like that, do you really need 20 in the blacks, given that there isn't a lot of perceptual difference between black and a dark grey, and generally those shades are compressed anyway?

and and accepted standards (like CoC).

I wasn't aware of an 'accepted standard' for CoC. There are several commonly used ones, but that's not a standard. As I remember, your CoC is 0.022mm for an APS-C image, which I can't make match any of the normal ones I know, The most common is I suppose d/1500, which would give 0.019mm. The Zeiss formula gives d/1730, which would give 0.016mm.

I'm not criticising your work, which is interesting and valuable. I just don't think you're in a position to claim that it is obviously more relevant than DxOmark's 'engineering DR'.