ORorbital-decay·Jan 16, 2022
It still seems we're talking about completely different things, or maybe you're trying to derail the conversation...
Both ML-based processing and tonemapping take place after demosaicing to produce the final JPEG. RAW capture is just your typical laplacian pyramid stacking/exposure fusion with motion compensation and sensor-specific calibration. That's the main contributing factor that somewhat compensates for the sensor size, not some kind of ML processing. If you think that ML is somehow involved here, I have zero idea where you are getting it from. The entire point of pre-demosaicing pipeline is to robustly converge the captured stack into a neutral scene-referred baseline with provably better SNR and precise photometric parameters. And that's what you get in a resulting DNG, not some neural tonemapped RGB image (which is not even crea
Read the rest
ted at this point).
Perhaps you should look at the pipeline yourself, here's the source code for a typical "computational imaging" camera app that produces stacked RAWs, for example. [1] And here's what Google Camera does [2] [3] (you can ignore everything that comes after demosaicing). Google's particular method is also partially implemented for arbitrary RAWs as third-party software, so you can stack any burst capture from your camera with the same method smartphones do. [4] (there's much less benefit for normal sized sensors, for obvious reasons)
[1] https://github.com/mirsadm/motioncam/
[2] https://static.googleusercontent.com/media/hdrplusdata.org/e... (PDF)
[3] https://static.googleusercontent.com/media/hdrplusdata.org/e... (PDF)
[4] https://github.com/martin-marek/hdr-plus-swift