sploo wrote in post #18116359
I think it was in an earlier post on this thread that John made a comment on the lines that Canon are out of the dog house - that is, the low ISO DR is sufficient that, whilst not quite matching that of rival sensors, it's now close enough that it shouldn't be a major factor in choosing a system (unless perhaps that was a very primary consideration... which means: buy a D810 and stick it on ISO 64).
ISO 64 should be possible with "normal" RAW headroom with dual-pixel RAW, if the headroom claims are true. Canon might not support it directly, but you can DIY with ETTR or maybe Magic Lantern might do something that can make it transparent to the user if they figure out the firmware.
True. I'm just a bit sad they weren't able to offer a quality 1.5x subsampling of the sensor, as that would have resulted in a tiny crop factor. IRC The 5D3 resolution (width) is exactly 3x1920, so I assume that was deliberate (to achieve a high quality 1920 result with much less moire than the 5D2).
That's computation-intensive. Now you're resampling, rather than binning like in the 5D3. Resampling can give the least aliasing, but 1.5->1 resampling is rough on details. They probably figured line skipping was not used in the 5D3, so that is no longer an option in the 5D series. Line skipping is the worst method, as far as aliasing goes. The best video from a Bayer sensor would come from a very high sensor resolution, and a heavy downsampling, with full color at every pixel and no aliasing, but that would require lots of processor power and battery power. Binning with overlapping source tiles could avoid aliasing, too, but that needs precise ratios of pixels. It's really a big struggle to resize to close numbers; the options are not ideal.
Indeed if they had been able to offer the 1.5x scaling on the 5D4 they could have even marketed it as having full frame and Super 35 modes. Ok, actually about a 1.09x and 1.64x crop respectively, but close enough for marketing
Like old Canons I guess?
I downloaded a set of D5 blackframes someone provided. I haven't done a direct comparison, but by sight and memory, the D5 seemed to be a bit worse than the 5D2 at ISO 100. That stuff shows up, too, at extremely high ISOs, when the pre-gain noise is very low and you downsample so that the random noise mostly disappears and only the banding noise is still strong. Banding noise and blotchy low-frequency noise do not disappear as fast when downsampling as random noise does. I suspect that even though the D5 outperforms the 5D4 and 1Dx at high ISOs in DxOMark, that in practice, if you pushed the Canons to ISO 3.28 million, and the final output was small 450*300web images processed by someone who knew how to handle the data with the least destruction, that the Canons may be a lot closer than DxO would suggest, or maybe even a little better. We don't see many tests like that, though. People tend to judge the usefulness of very high ISOs by 100% crops or large images sizes, where they are likely to be disappointing. Yet, people ooh and aah at small web images at medium-high ISOs, which are not very challenging at all if the the processor knows how noise, filtering, and resizing work.