Messages

For the shadows the ISO 100 lines are being discarded from the composite raw, which reduces the light-capturing surface area of the sensor in half.

Thanks Horshack-- funny how the obvious sometimes escapes me. Although, b/c they're interpolating between the two ISO exposures, you can set up the algorithm to only use the ISO 1600 file where read noise trumps shot noise contributions significantly. But I see what you're saying -- when you do that, you pay a cost due to shot noise contributions (vs. exposing the whole image at ISO 1600, for example).

But typical 'engineering' DR calculations using SNR of 1 as the lowest signal don't consider shot noise, correct? Therefore, one could still say that the engineering DR is extended to FWC/read noise @ISO 1600, which is ~14EV. If you're using a higher SNR cutoff of, say, 20, then the DR is much more modest (considering the increase in shot noise contributions you mention).

I read through the PDF and I saw the section on deep shadows where you only use the ISO 1600 data, indicating "corresponding data from ISO 100 will be just noise". The ISO 100 lines will have more read noise and tossing them out will improve DR but at the same time will reduce photons collected by half, reducing the shot-noise SNR of those shadows from 22.93:1 to 16.21:1

Hi horshack-- why would you 'reduce photons collected by half'? The exposure at the sensor plane is not changed; there is only one exposure. The ML hack is simply changing what you do with that exposure data, & so should not affect the SNR save for actually increasing it in shadows by reducing the effects of downstream noise (after ISO amplifier) in the ISO 1600 exposure.

The focal plane exposure (in terms of shutter speed/aperture) remains the same for both ISO 100 & ISO 1600 exposures; it's not like two separate exposure are being made (where the ISO 1600 exposure would have 4 stops less exposure which, yes, would increase the effects of shot noise).

Let me know if I'm misunderstanding you.

Wonderful work, & great white paper!

The funny thing about all this is that this is all a workaround largely b/c of the ADCs being on a separate chip from the image sensor (as surmised by ChipWorks). Essentially: your technique is trying to recover the dynamic range of the CMOS sensor itself, not the sensor & the entire signal processing chain. This'd all be pretty unnecessary on a D600/D800, where there's not much difference between the two (although the higher ISO exposure could help for very very deep shadows that might otherwise suffer from quantization error at ISO 100). The actual CMOS sensors on Canons have a similar DR to that of the best Sony sensors, if we believe the numbers from DXO/Sensorgen.info (though, this doesn't take into account FPN; I'm not sure exactly where in the chain FPN is introduced).

Do you know what the difference between the top & left black bars is? I've seen this & often wondered.