Kittyk
Offline
• • • •
Upload & Sell: Off
|
 p.4 #18 · Would you prefer D800 with or without AA filter ? | |
theSuede wrote:
Nope, not even close. Not by a country mile, and not by a light-year either. It's easier to calculate a trajectory for a moonlanding than what it is to reverse-engineer a sparsely sampled geometry.
What you're trying to do is fill a sparsely sampled unknown geometrical surface without any sample support - this is as close to impossible as you can get on any patch area larger than 10x10 samples. You need at least 100x100 pixels to get close to an accurate pattern recognition.
*And you need a pattern matching scheme run by an AI to get away from the "brute force" method of simply iterating until you find a matching sample.
*And say you wrote this piece of software to absolute perfection in optimization.
*And say that this patch is centered in the image, so that you don't have to run multiple CA scenarios
*And say that you have an optimized database of patterns for the recognition software to use
You would still end up with about n^3*(10000*15)*(10000*15) ops and about half as many memops. A bit more than a second per pixel, given the absolutely best of circumstances. Now run this over a few million pixels (add in inter-pixel pattern recognition gains), and you have a processing time of about week per MP of image area, in the fastest computers available today. You can make some safe assumptions and conscious quality restraints, and get down to an hour per photo, I'd think. But then you'd have some SERIOUS quality restrictions.
Let's say that I need an accurate stone count and stone layout in "this field". The stones are 0.8m diameter, and the field is 100x100 meters. I give you 5-10,000 different patterns in which the stones can be laid out on in the field (none of them may be correct, and one hundered of them may be almost correct - at the same time), and a map containing a "stone percent count" on every 2m gridpoint. I need the stone count, AND pattern for every 1m gridpoint.
.............
It is quite simply put - impossible.
What you can do is to minimize the damage, and ask permission to just give an average stone count per 10x10m grid - or something like that. Which is just what de-moire software filter do. They lower your resolution by a factor of ten or more.
How much this hurts the image depends on the image. In most very little (5% reality loss?), in some you just can't undo the damage in any way except manual inpainting (100% reality loss)....Show more →
try capture one, you will be surprised. there are many other methods, like using sort of fourier transformation to approximate the moire frequency or even brute force. also keep in mind that AI is replaced by photographer with Logitech G series keyboard playing with sharpening, clarity, moire amount, moire pattern sliders with 80-200 levels of adjustment each until it matches desired input. And i can assure you, that i do not lose the resolution more then 20% by that (wild guess) and i lose it only in cases, and only in areas, where it is problematic. In all other areas, i get more then i would have with AA filter. I also stand behind, that after this smart moire correction, i most probably still have more details then with AA filter on. Or at least with as strong AA filter which would fully remove the moire in first place.
Also keep in mind, that given the scene is same, you actually have less likely moire issues as you up the resolution. And as you up the resolution, it is less likely going to negatively affect the final print.
60 Mpixels means, you get fine fabric details on full body portrait*. You get sometimes moire in say, 20cm² area, depending on how the fabric is twisted. Easy. Sure you have bazilions of combinations but in real world you usually get small area with pretty repetitive pattern which is actually quite easy to fix.
36Mpixels is of course not as good, but will be better then moire of D700 without AA filter.
So it is more like, hey, we got 36Mpix, now WE CAN remove the AA filter because it will in real world much less disturb then if we would do it on D3x.
I am not mathematician, but i am doing it almost every day. And if you want to test me, give me moire loaded RAW to fix, and then compare resolution.
*i am in portrait business, so more details then eye lashes i don't need. Anything finer then eye lashes (fabric, mainly) is not interesting for me.
|
). It's inherent to discretely sampling a continuous signal. 