skibum5 wrote:
that said Iliah Borg often knows what he is talking about, so maybe his magic is not magic but real, becomign a little dubious, but it would be fantastic if true, he seems to claim the full RAW show solid clipped colors in highlights where he claims you can see well distinguished values and he claims there seems to be no weird histograms holes or anything, but I agree it does seem all a bit strange and seems a little hard to believe....
I now think that Iliah is right.
And yesterday I likely was entirely wrong, despite my confidence and conviction.
I have completely changed my mind. The main (complete) image is considerably clipped (14% of the area upon the shown stats) but the extra subimage isn\'t at all: both saturate at the same point, but the exposure is obviously not the same (I must admit that this was just what the histograms in the dpreview thread/RawDigger web already shown. I believe that DPRAW extra secondary RAW image is perfectly handled by that application and the shown data is correct).
Anyway there is something not right here.
If all of this is as true as it seems, Canon is effectively wasting 1 EV of dynamic range. They could have fixed this, e.g.:
Compensating the RAW values in the combined image. Just like ISO 160 is obtained from ISO 200 multiplied by 0.8 in many cameras, this time they could have used 0.5 to move the histogram back to the left. I think that wouldn\'t damage the data, only remove a bit of noise from the shadows (well, that would also yield a darker image... and maybe Canon didn\'t want to tag the combined ISO 100 subpixels, to honor the exposure meters, as an ISO 50 for the whole image).
Storing an additional 15th bit in the RAW file, but this could break the file specs and/or turn crazy the converters. And I think it is not really required, because the sensor doesn\'t goes beyond 14 bits of DR at the pixel level (the properly combined subframes, without clipping, would yield about 13.6 EV at pixel level).
Combining images adds their signals, but the noise gets combined by the square root of their squares. So to yield the read noise of 2.5 ADU for the main image that we already have measured on average, we should expect a read noise of 1.77 ADU in each subimage:
2.50 = sqrt(1.77^2 + 1.77^2)
This implies a DR of log2((16383-512)/(1.77)) = 13.1 EV for each separate subimage at the pixel level (14.1 EV DXO-normalized, compared with the 13.6 for the main combined image). That is, thanks to not clipping the highlights, and even having only half the signal, the extra subimage likely outperforms in DR the main combined image by about 0.5 EV!. But using only the subimage I assume that sharpness would decrease. It would be interesting to compare both at 100%.
And also it would be interesting to experimentally measure the read noise of the extra subimage, but DCRAW can\'t extract it (I don\'t know if RawDigger shows the optical black area, I confess that I haven\'t tried it). If it finally is not 1.77 that would break my reasoning and could hint something else.
Another interesting thing to look at would be a ISO 50 RAW file from 5D4.
If this negative feature is totally confirmed coming for free, it there could be also commercial or marketing reasons behind the Canon decision. But it would be not easy at all to assimilate, because users value a lot each bit of DR... and this is a full bit!
skibum5 wrote:
that said Iliah Borg often knows what he is talking about, so maybe his magic is not magic but real, becomign a little dubious, but it would be fantastic if true, he seems to claim the full RAW show solid clipped colors in highlights where he claims you can see well distinguished values and he claims there seems to be no weird histograms holes or anything, but I agree it does seem all a bit strange and seems a little hard to believe....
I now think that Iliah is right.
And yesterday I likely was entirely wrong, despite my confidence and conviction.
I have completely changed my mind. The main (complete) image is considerably clipped (14% of the area upon the shown stats) but the extra subimage isn\'t at all: both saturate at the same point, but the exposure is obviously not the same (I must admit that this was just what the histograms in the dpreview thread/RawDigger web already shown. I believe that DPRAW extra secondary RAW image is perfectly handled by that application and the shown data is correct).
Anyway there is something not right here.
If all of this is as true as it seems, Canon is effectively wasting 1 EV of dynamic range. They could have fixed this, e.g.:
Compensating the RAW values in the combined image. Just like ISO 160 is obtained from ISO 200 multiplied by 0.8 in many cameras, this time they could have used 0.5 to move the histogram back to the left. I think that wouldn\'t damage the data, only remove a bit of noise from the shadows (well, that would also yield a darker image... and maybe Canon didn\'t want to tag the combined ISO 100 subpixels, to honor the exposure meters, as an ISO 50 for the whole image).
Storing an additional 15th bit in the RAW file, but this could break the file specs and/or turn crazy the converters. And I think it is not really required, because the sensor doesn\'t goes beyond 14 bits of DR at the pixel level (the properly combined subframes, without clipping, would yield about 13.6 EV at pixel level).
Combining images adds their signals, but the noise gets combined by the square root of their squares. So to yield the read noise of 2.5 ADU for the main image that we already have measured on average, we should expect a read noise of 1.77 ADU in each subimage:
2.50 = sqrt(1.77^2 + 1.77^2)
This implies a DR of log2((16383-512)/(1.77)) = 13.1 EV for each separate subimage at the pixel level (14.1 EV DXO-normalized, compared with the 13.6 for the main combined image). That is, thanks to not clipping the highlights, and even having only half the signal, the extra subimage likely outperforms in DR the main combined image by about 0.5 EV!. But using only the subimage I assume that sharpness would decrease. It would be interesting to compare both at 100%.
And also it would be interesting to experimentally measure the read noise of the extra subimage, but DCRAW can\'t extract it (I don\'t know if RawDigger shows the optical black area, I confess that I haven\'t tried it). If it finally is not 1.77 that would break my reasoning and could hint something else.
Another interesting thing to look at would be a ISO 50 RAW file from 5D4.
If this negative feature is totally confirmed coming for free, it there could be also commercial or marketing reasons behind the Canon decision. But it would be not easy at all to assimilate, because users value a lot each bit of DR... and this is a full bit!
Sep 01, 2016 at 01:54 PM
Previous versions of cgarcia's message #13704880 « Canon 5D4 dynamic range analyzed from RAWs »
