Pentax dynamic (handheld) pixelshift


(nosle) #1

Dpreview have done some tests and I’ve found some artefacts suggesting that dynamic pixelshift might indeed be using non-debayered data. At least that’s my conclusion after seing these familiar pixel level artefacts.

Perhaps they aren’t a result of raw pixel merging but they look like it to me. Note the review claims it’s debayered but they don’t reference any source.

edit: a better link is the following https://www.dpreview.com/articles/7242776989/pentax-k-1-mark-ii-dynamic-pixel-shift-mode


(Ingo Weyrich) #2

I agree that it looks exactly like the familiar pixel level artefacts.
I just downloaded the handheld mode raw files. Currently they don’t have the PS icon bottom right on the thumb in rt file browser, because it’s a new mode. That’s easy to fix. I will include it into my next push to pixelshift_fixes branch. Though the pixelshift algorithm won’t give good results for this files. It’s about the same as using the non handheld PS mode without a tripod.


(Andrew) #3

guys, can I see if I understand this…
If you are confident you have data sets where there is exactly one pixel shift between each, then you can make special use of this knowledge, produce more accurate colours than normal, and the process is then quite different from a normal debayering/demosaicing.

But if it’s handheld and therefore random shifts, you have to demosaic each data set separately then do an alignment + averaging process.

Is that about right?


(Ingo Weyrich) #4

Absolutely, that’s what RT does for the well defined standard PS files. With the extention, that RT is able to detect motion and use the demosaiced frame of your choice for regions with motion.

If the random shifts are known (maybe Pentax stores them in metadata for the new handheld mode) it could be possible to align the images before demosaicing, but at least currenty, even if we had this data, I 'm not willing to waste a minute for implementing that.


(Graeme W. Gill) #5

Couldn’t you demosaic, find best alignment, then use that to align before redoing the demosaicing ?


(Martin Godwyn) #6

Am I reading this right?

“Though the pixelshift algorithm won’t give good results for this files. It’s about the same as using the non handheld PS mode without a tripod.”

Given that regular (non-hand held) PS is useless without a tripod, this makes it sounds like the new Dynamic PS is equally useless without a tripod, which of course defeats its very purpose. Is that what you’re saying?

My impression (well, my guess, really) was that the new Dynamic PS essentially combines the IS with the PS. That is, employing IS to keeping the sensor essentially still aside from shifting the sensor in one pixel increments to produce the PS. If that is the case then the data sets would be well defined, no? Have I got that wrong? If it were just processing four random shifts then it really isn’t “pixel” shift anymore.


(Alberto) #7

this sounds like something worth trying at least (without having any idea of how this new dynamic ps mode works at all though – but it seems a neat trick on paper :wink: