Pentax 645z vs Fuji GFX

[Jim Kasson]

I can't believe you have so much difficulty understanding the problem with your post. The post only addresses one part of your argument as to why the magical microlenses are so so revolutionary in achieving superior levels of sharpness with the same lens, namely the size of the microlenses. The second part, which I expected to see discussed but never was, is WHY smaller microlenses make a difference in sharpness that cannot be achieved with the "regular" microlenses and highly skilled sharpening techniques. If you had to explain to a group of people who are not engineers why the Fuji microlenses are so significant, would you just trot out Brian Dube to deliver his scientific paper as it was written?
What makes me so skeptical about the magical microlenses thing is that I actually tested the GFX against another camera with the same sensor but with "regular" microlenses. Yes, if you just took GFX files into LR and applied the default sharpening in LR, they looked very sharp. However, they couldn't handle much additional sharpening without becoming crunchy. The X1D files did not look as sharp with the default sharpening in LR, but as soon as the settings were changed to achieve optimal sharpening in LR, the X1D files looked sharper. The same thing was observed by turning off LR sharpening, bringing the files into PS and using the same Focus Magic settings. Moreover, there are some really smart people working at Pentax, Phase and Hasselblad, and none of them reengineered the microlenses. Were they all just asleep at the switch? Or is there a tradeoff in using the magical microlenses that they considered a poor design choice? Soon enough we will know, as the next iteration of the 44x33 medium format Sony sensor will be released next year. We shall see if they all use the same size microlenses.
So, if I understand what you are saying about the HC lenses, the comparison I suggested wouldn't work because the HC lenses suck so bad. The magic microlenses only show their stuff with certain lenses but not others?
I did see your tests with the HC lenses on the GFX. I agree that the leaves on the trees at a distance looked bad. But to be candid, the leaves on the trees in all your tests look really bad to me. Last weekend, I photographed a tree at a long distance with the HC 210 and 300mm lenses and enlarged the files to 300% on the screen and the leaves did not look nearly as bad.

I said above that there is no magic in the small microlenses. They are a win for sharpness, as they are a bit -- but only a bit -- closer to the ideal point sensor. But they lower overall quantum efficiency and increase the probability of aliasing. Indeed, they are in some ways a return to the old days, before microlenses when fill factors were much less than 100%. It's not clear to me that they are a better overall tradeoff than having 100% fill factor. But then again, I am on record as being for AA filters on MF cameras, and that is not a popular position.

I addressed the sharpening question above, using the D800/D800E as examples.

As to why the GFX files get crunchy on sharpening, it may be that they don't need it. It may also be that it's hard to get a deconvolution kernel for a lens with so small an effective fill factor since the GFX is an outlier in that regard. Where did you get your kernel when you did your tests? In general, deconvolution sharpening works well when you have the right kernel and when the SNR is high to start out with, since sharpening adversely affects visible noise.

I did not say the HC lenses suck. On a physically-larger sensor of around 40MP, like the ones for which they were designed, they are pretty good. The sharpness of the file is the product of the MTFs of the sensor and the lens. If the lens is not delivering the sharpness, having a sharp sensor won't help much. If you think of it in terms of blur circles, the two add approximately as the square root of the sum of the squares, so the bigger circle takes over pretty soon. I never owned the 210 and 300 HC lenses.

You will note in all the discussion above, I never said that small microlenses are overall better than large ones, which is an opinion that it seems that you think I have.

Jim

[Bart_van_der_Wolf]

I said above that there is no magic in the small microlenses. They are a win for sharpness, as they are a bit -- but only a bit -- closer to the ideal point sensor. But they lower overall quantum efficiency and increase the probability of aliasing. Indeed, they are in some ways a return to the old days, before microlenses when fill factors were much less than 100%. It's not clear to me that they are a better overall tradeoff than having 100% fill factor. But then again, I am on record as being for AA filters on MF cameras, and that is not a popular position.

I agree, it's not popular (with traditional MF camera-back users, because they usually have no other choice). But from a Digital Signal Processing (DSP) point of view, a 100% fill-factor and/or a proper Optical Low-Pass Filter (OLPF) does reduce the tendency for generating aliasing artifacts (and problematic demosaicing).

As to why the GFX files get crunchy on sharpening, it may be that they don't need it. It may also be that it's hard to get a deconvolution kernel for a lens with so small an effective fill factor since the GFX is an outlier in that regard.

The latter is a more likely explanation IMHO. It requires a differently shaped deconvolution filter.

If you have the right tools (and I know it was a rhetorical question), then a reduced fill-factor deconvolution filter can be gotten by creating one, e,g, with my online PSF-generator tool here. Finding a tool that uses it as input is somewhat harder ...

Cheers,
Bart

[Jim Kasson]

I said above that there is no magic in the small microlenses. They are a win for sharpness, as they are a bit -- but only a bit -- closer to the ideal point sensor. But they lower overall quantum efficiency and increase the probability of aliasing. Indeed, they are in some ways a return to the old days, before microlenses when fill factors were much less than 100%. It's not clear to me that they are a better overall tradeoff than having 100% fill factor. But then again, I am on record as being for AA filters on MF cameras, and that is not a popular position.

These don't address the QQE issue, but here are a couple of visual indications of both the improved above-Nyquist contrast available from smaller microlenses and the increased aliasing. They are simulated images of a 60-spoke raised-sinusoid Siemens Star with fill factor starting at 10% and increasing in steps of 10% to a maximum of 90%, imaged on a sensor with a 5.3 um pitch through a perfectly-focused diffraction-limited lens set at f/2.8. You can also see the effect on false color.

First, with AHD demosaicing:



and then with bilinear interpolation demosaicing:



If you've done any imaging with Zeiss Siemens Star targets and real cameras, you will probably find the contrast and shot-to-shot stability of the above images better than you can obtain in real life, but if gives an indication.

As to the interaction of small FF microlenses -- and AA filters -- with deconvolution sharpening, I'm still working on how to test that fairly.

Jim

[eronald]

I said above that there is no magic in the small microlenses. They are a win for sharpness, as they are a bit -- but only a bit -- closer to the ideal point sensor. But they lower overall quantum efficiency and increase the probability of aliasing. Indeed, they are in some ways a return to the old days, before microlenses when fill factors were much less than 100%. It's not clear to me that they are a better overall tradeoff than having 100% fill factor. But then again, I am on record as being for AA filters on MF cameras, and that is not a popular position.

I addressed the sharpening question above, using the D800/D800E as examples.

As to why the GFX files get crunchy on sharpening, it may be that they don't need it. It may also be that it's hard to get a deconvolution kernel for a lens with so small an effective fill factor since the GFX is an outlier in that regard. Where did you get your kernel when you did your tests? In general, deconvolution sharpening works well when you have the right kernel and when the SNR is high to start out with, since sharpening adversely affects visible noise.

I did not say the HC lenses suck. On a physically-larger sensor of around 40MP, like the ones for which they were designed, they are pretty good. The sharpness of the file is the product of the MTFs of the sensor and the lens. If the lens is not delivering the sharpness, having a sharp sensor won't help much. If you think of it in terms of blur circles, the two add approximately as the square root of the sum of the squares, so the bigger circle takes over pretty soon. I never owned the 210 and 300 HC lenses.

You will note in all the discussion above, I never said that small microlenses are overall better than large ones, which is an opinion that it seems that you think I have.

Jim

Jim,


maybe with a small fill factor and a good lens you'll get more aliasing, eg. an mtf that is non zero well past the Shannon limit?
i think there is only so much understanding one can obtain from the easily computed mtf curve. Also, note that if the optics are decent for green, they are going to be more than decent for r,b which have half the limit frequency, and there the aliasing from the sensor is going to hit you in the face, translating into a false-color explosion. Here the values will also be subject to massive numerical aliasing - sampling errors- and then applying a deconvolution will blow up this numerical noise, with artefacts all over the place. I'm just making all of this up as I go along of course, but you must admit it sounds nice and explains why we see fewer sensors with small fill factors and why old MF sensors were called "aliasing machines". I'm sure you have simulation code somewhere which will disprove all my vague intuitions unless by some unlikely accident they happen to be true.

Edmund

[Jim Kasson]

Jim,


maybe with a small fill factor and a good lens you'll get more aliasing, eg. an mtf that is non zero well past the Shannon limit?
i think there is only so much understanding one can obtain from the easily computed mtf curve. Also, note that if the optics are decent for green, they are going to be more than decent for r,b which have half the limit frequency, and there the aliasing from the sensor is going to hit you in the face, translating into a false-color explosion. Here the values will also be subject to massive numerical aliasing - sampling errors- and then applying a deconvolution will blow up this numerical noise, with artefacts all over the place. I'm just making all of this up as I go along of course, but you must admit it sounds nice and explains why we see fewer sensors with small fill factors and why old MF sensors were called "aliasing machines". I'm sure you have simulation code somewhere which will disprove all my vague intuitions unless by some unlikely accident they happen to be true.

You guessed it. I've posted some sim results today that show how diffraction and fill factor interact.

http://blog.kasson.com/the-last-word/siemens-star-diffraction-effects/

I ran experiments today -- not posted yet -- where I ran a natural image (foliage) through the sim at 10%. 60%. 100% fill factors with diffraction-limited lenses from f/2 to f/32. I am amazed at how good the f/2 10% FF image looks, and it must have aliasing up the wazoo.

Jim

[eronald]

You guessed it. I've posted some sim results today that show how diffraction and fill factor interact.

http://blog.kasson.com/the-last-word/siemens-star-diffraction-effects/

I ran experiments today -- not posted yet -- where I ran a natural image (foliage) through the sim at 10%. 60%. 100% fill factors with diffraction-limited lenses from f/2 to f/32. I am amazed at how good the f/2 10% FF image looks, and it must have aliasing up the wazoo.

Jim

I think false detail is fully acceptable - maybe even desirable - on natural textures, foliage, bark, stone, probably because it preserves the self-similarity measures by which our perceptive processes recognize texture. As defined by another IBM fellow, Benoit Mandelbrot who incidentally was a graduate of, and on a sabbatical year at Ecole Polytechnique when I was blundering my way through my PhD. This actually would be worth looking at at mathematically or at least by means of your simulation.

So, the images from the old 22Mp backs are pleasing because they really do *look* sharp - I am willing to bet that if we looked at the sensor MTF here we would see a substantial residual value past the Shannon limit corresponding to the fictitious detail generated by aliasing.

Anyway, my interpretation of your very nice binary stars is that now that fill factor has gone up cameras mostly don't need AA filters, which of course is why many manufacturers proudly advertise their absence as a feature

As we both know the computations you have been doing can be found in the remote sensing literature where it has been usual for some time to compute the sensor MTF.

At this point I'll bow out, this forum has seen enough of my pompous self for a few months, and anyway in my old age I've replaced photography with painting.

All the best to you, Jim.

Edmund

[Jim Kasson]

I think false detail is fully acceptable - maybe even desirable - on natural textures, foliage, bark, stone, probably because it preserves the self-similarity measures by which our perceptive processes recognize texture.

Here's a test that bears that out, at least for foliage:

http://blog.kasson.com/the-last-word/simulated-foliage-shots-vs-fill-factor/

In this test, WRT fill factor, there's less there than initially meets the eye.

Jim

[eronald]

Here's a test that bears that out, at least for foliage:

http://blog.kasson.com/the-last-word/simulated-foliage-shots-vs-fill-factor/

In this test, WRT fill factor, there's less there than initially meets the eye.

Jim

I have an idea for a paper if you're interested, we could rope in Norman Koren from Imatest 

Your simulator, his MTF software, my bright personality and writing

Edmund