I've found some old DNGs on the Apertus website (October 2013), which had strong FPN, so I took one of them and tried to fix it.
The settings were unknown (I believe they were not shot at the base ISO), and none of the dark frames from the previous experiment worked. So I've tried to estimate the FPN and row noise from scratch, like this:
- used a bilateral filter to remove low-frequency details and strong edges
- used two strong line filters (one horizontal, the other vertical) to detect the FPN from the remaining high-frequency details
The result is not perfect (the correction introduced some artifacts, and there is still some FPN left), but the improvement is noticeable. If the Beta hardware will have lower FPN than the Alpha (as Sebastian hinted
here), the remaining noise should be much easier to fix in software.
The image was overexposed by about 2 stops, but the blue channel was not clipped at all. It was too much for ufraw's highlight recovery algorithm, so I've applied my own (I've talked about it
here).
The image also showed artifacts (
like these), so I've used the VNG4 algorithm for demosaicing to avoid them.
Black level was also unknown, so I've also guessed it (0.1 percentile).
Left to right:
- original
- after fixing FPN
- after fixing FPN and recovering highlights
Developed with UFRaw + enfuse (DNG developed at roughly 0, +2 and +4 EV).
Full-size images, DNG files and UFRaw settings:
https://www.dropbox.com/sh/j8xalnjw5cn7w8t/AAAi43GM33R2Q6jHP4JdwDNqa?dl=0Feel free to push them and show your results, but don't expect to be as clean as a dual ISO image
Any chance of plotting the data of the CMV20000 on that chart?
It's in my first post from this thread (only the base ISO, from the datasheet):
The base ISO seems on par with 5D3's ISO 400; the question is how much it will catch up at higher ISOs (can't tell from datasheet). It seems slower (30fps at full resolution according to
this). Sebastian, do you know if there's a V2 for this sensor as well?
