Pixel shift resources

[Lou Jost]

I was referring to Olympus' theoretical expectation, as evidenced by their use of 50Mp jpgs from 80Mp RAW files. Another interesting clue to Olympus' expectations about this shifting process is that they prevent us from using apertures smaller than f/8 for pixel shifting.

There is a nice accessible theoretical discussion of the 4-image pixel shift of Sony and Pentax here:
https://www.wrotniak.net/photo/m43/em1.2-hires.html

Somewhere I have seen another analysis but I can't find it now.

Edit to add another informative practical link:
https://www.stevehuffphoto.com/2016/08/ ... ad-nichol/

[chris_ma]

I find it's hard to quantify exactly because of the topic of aliasing/moire.

there's the rule of thumb floating around that a bayer sensor has about 70% of the native sensor resolution (ie a monochrome sensor that would use a RGB filter wheel), but I found it really depends.

from what I've read modern debayer algorithms actually try to recognise the subject structure and adapt their processing based on that (even locally in the same RAW file) - that is why lightroom often has such clean results compared to simpler debayering (dcraw is very useful if you like to experiment with these things).

however, given a sharp enough lens and enough complexity in the subject, we can run into moire patterns from the debayering and then resolution drops a lot. compare different patterns on the dpreview scene:
https://www.dpreview.com/articles/35281 ... -published

if you look at the text upper center, the Phase One 100MP clearly has highest resolution. The Panasonic, while soft, drops a lot more then 40% when switching to single exposure.

And if you look at the color circle pattern suddenly the Phase One is nor really better then the others anymore, and the Panasonic in single exposure again looses a lot of resolution (and produces tons of false detail).

and this is with lenses shot at F5.6 and F8 and test patterns that look like they are slightly soft as well.

So it seems to me that pixel shift is most useful on small patterns that introduce moire and less on larger image detail or clean lines etc.

chris

[Lou Jost]

"there's the rule of thumb floating around that a bayer sensor has about 70% of the native sensor resolution (ie a monochrome sensor that would use a RGB filter wheel), but I found it really depends."

Yes, it really depends. The first link in my last comment dissects the possibilities and quantifies the range of improvement.

pixel shift is most useful on small patterns that introduce moire and less on larger image detail or clean lines etc.

That's certainly right for the four-frame pixel shift algorithms, but the eight-frame algorithms do more than eliminate moire, they increase real resolution, so they should increase the accuracy of edge detail (though they may reduce the contrast at the same time).

[chris_ma]

That's certainly right for the four-frame pixel shift algorithms, but the eight-frame algorithms do more than eliminate moire, they increase real resolution, so they should increase the accuracy of edge detail (though they may reduce the contrast at the same time).

agreed, this can be observed in this area for example:



the pixel shift clearly gives less moire and more resolution, but looking at the single vertical line it also has way softer edges then a single 100MP bayer picture

[rjlittlefield]

I was referring to Olympus' theoretical expectation, as evidenced by their use of 50Mp jpgs from 80Mp RAW files.

From my standpoint, that doesn't look like a "theoretical expectation". It looks more like some compromise that stems from the marketing question "What would our customers be happiest with?"

I can easily imagine a decision process along the lines of "Well, anybody who wants a JPEG at all probably wants one that looks pretty good without further manipulation. We know that our 80 Mp raw files have low contrast for fine detail, and we can't sharpen them up to look good at actual pixels, but we can downsample and get a 50 Mp that looks pretty good even at 100%, so let's just do that. If anybody really wants the last bit of detail, let them dig it out of the raws."

Another interesting clue to Olympus' expectations about this shifting process is that they prevent us from using apertures smaller than f/8 for pixel shifting.

Here's a simple calculation to explain that. With pixel-shift sampling, the Olympus OM-D E-M1X has an effective pixel size of 1.68 microns (sensor pitch 3.36 microns, divided by 2). At 550 nm and f/8, the diffraction limited cutoff frequency is 4.4 microns per cycle, which can be resolved at the Nyquist limit by 2.2 micron pixels. So at f/8, the lens is already under-resolving the pixel-shifted sensor. Allowing smaller apertures would only result in actual pixel views looking even more blurred.

--Rik