skibum5
Offline
• • • • • •
Upload & Sell: Off
|
 p.4 #16 · Canon 5D4 dynamic range analyzed from RAWs | |
from canon learning:
"Before a picture is taken, for focus purposes, light is gathered from the A and B pixel areas separately, and read as two independent signals.
Each light-sensitive side of a pixel can be independently controlled.
But when an actual still image (or video frame) is recorded an instant later, the two separate signals from the A and B sides of each pixel are combined, to generate one single brightness signal from each pixel. Thus, each pixel gets essentially the same full photo diode area that a convention sensor (without Dual Pixel CMOS AF technology) would have.
In effect, Dual Pixel RAW images have a full “layer” — figuratively speaking — of digital information from each pixel, using the conventional combined A and B output from all the pixels on the imaging sensor. But in addition, they add what we can think of as parallax information from just the A-side of each pixel."
So the sensor seems to be able to combine in HW and send that out and also be fast enough to send out A side only information. For whatever reason it seems like when the hardware adds the two sides everything goes up 1 stop and the top stop gets clipped 1 stop compared to just the A side alone (and I'd assume the non-accessible B side alone). Maybe it has only 14 bit HW and it adds them in 14 bits which overflows one stop and then it divides back down by 2 or something.
It seems like instead of going for this micro-shifting and slight bokeh and flare change stuff they could have instead of sending out A channel, taken A/2 + B/2 and written that to the extra channel and then from that you could get the extra stop of highlights but taken from combined channels so no potentially weird side to side lighting effect shifts for the top stop and then you'd have your ability to get your full extra stop without any bad issues possible at all. But maybe that is too much math to be put on the sensor.
It also seems like it seems like their sensor+new ADC on chip system actually can deliver as much DR as Nikon can manage at ISO100 or in that range but by going with DualPixel design for AF for video and using 14bit HW they are losing a stop and it seems like perhaps all dualpixel bodies could have had one more stop of DR had they been traditional AF cameras only ? Or if they simply decide to care more about DR than this minor focus shift, bokeh shift stuff.... honestly I sort of think it would've been far more worth it to just deliver the extra stop without compromise than this bokeh, flare, focus micro shift stuff which I'm not sure will end up being used all that much once the novelty wear off. I mean I'm sure it could be useful at times but....
Also almost sounds like if they could just stick something that could add two 14 bit values and then divide and just retain one extra bit before dividing they could have best of both worlds, but for one reason or another that is not done yet??
Perhaps one can use the A channel to truly extend the DR another full stop, I just wonder if the slightly directly pickup of light to A channel side will create too many weird issues in the final result though. Would really be a shame if that is the case. I also wonder if it would be a constant correction factor and you could just apply a slight increasing brightness shift horizontally across the screen, but it might depend upon where light sources were and be non-constant, but not sure, it might be possible to correct it uniformly in software, sure it costs a touch of SNR yet again, now for the second time, but then again as I said before who the heck cares about SNR for the top stop at ISO100! It's so good up there I doubt you'd notice any noise added. Maybe it can work?
Hmm sounds almost like they could have delivered a camera that totally matched D810 for DR at low ISO but decided to instead give it DualPixelAF at the cost of that stop (which I guess would also be a full stop DR lost at any ISO too even very high ISO? where DR is really at a premium) or since they maybe could have found a way to still deliver it at the cost of 60MP file size had they simply decided to focus on DR over this weird bokeh/focus/flare stuff (although maybe what would be need to be done to save the top stop would heat the sensor too much or slow it down too much or something?).
Not sure at what stage what is done, so not sure if it's the sensor or even DIGIC where the clip happens, if DIGIG then it's a shame they didn't make a DIGIC 6++ that could just have a new instruction with a single 15 bit value buffer or something for doing a simply 'long' math instruction that adds, keeps extra bit divides by 2, returns 14bit results. but maybe that's hard to pull off.
COuld be wrong though, very tired and brain is asleep. And something about what I wrote above feels a bit off and not entirely sensible.
more:
"In answer to the inevitable question about using third-party RAW processing software from other companies, Canon has no idea whether any of these software programs will have the ability to leverage the Dual Pixel RAW technology and add processing options beyond ordinary RAW image adjustments. It’s up to each independent company making their own RAW process software to take RAW files from Canon cameras, reverse-engineer what’s needed to process them into finished images, and incorporate this into updated versions of software that are compatible with RAW files from the latest cameras."
wonder why they make them reverse engineer it all, seems like stabbing themselves in the back
Edited on Sep 01, 2016 at 09:30 PM · View previous versions
|
