Absurd DxO D800E results: Are low pass filters a joke?

[eronald]

Jim, Bart,

  This is an interesting topic. I would like to take this further. In particular, I would like to understand what MTF curves for a real-world AA filter looks like.
  Unfortunately I am due for a minor medical procedure later today, and will probably not have the necessary clarity of thought for serious discussion for a few days - my apologies and I hope to revisit the topic soon. Please do not  take my lack of response in this thread as an indication of disinterest.

Edmund

[Fine_Art]

Using the 144 cycle Siemen's star chart you may get an idea of what frequencies are attenuated the most. It would be nice to have one made of ground glass with black paint and a backlight. That would give the contrast needed for a better effect. A printed 400:1 contrast does not have the proper MTF attenuation.

[Bart_van_der_Wolf]

Jim, Bart,

  This is an interesting topic. I would like to take this further. In particular, I would like to understand what MTF curves for a real-world AA filter looks like.
  Unfortunately I am due for a minor medical procedure later today, and will probably not have the necessary clarity of thought for serious discussion for a few days - my apologies and I hope to revisit the topic soon. Please do not  take my lack of response in this thread as an indication of disinterest.

Hi Edmund,

No worries, getting well has a higher priority.

Cheers,
Bart

[Bart_van_der_Wolf]

Using the 144 cycle Siemen's star chart you may get an idea of what frequencies are attenuated the most. It would be nice to have one made of ground glass with black paint and a backlight. That would give the contrast needed for a better effect. A printed 400:1 contrast does not have the proper MTF attenuation.

Hi Arthur,

There are commercial Chrome-on-glass versions available for lens testing. However, they are not suitable for sampled / digital camera imaging, because the sharp edges add high spatial frequencies that will lead to aliasing artifacts. Also, the high contrast will not be needed, and can even distort the read-out (due to glare), and cause trouble in subsequent processing (more sensitive to built-in sharpening, causing overshoots and clipping).

These black/white targets are not suited for digital imaging (they are for analog imaging, vibration testing, or inspection of the lens on an optical bench). That's why I invented the sinusoidal version more than a decade ago (it was a 60 cycle version, because it was less demanding to print with the technology of those days).

Cheers,
Bart

[Mr. Capp]

Now the tests are all shot wide open correct?
I never shoot wide open, if ever above f4 mostly 5.6-11.
Would any diffraction still equalize the bodies at smaller apertures.
Before I bought the d800 everyone said "the difference is so minor" "sharpening will minimize any difference"
These tests seem to indicate a pretty huge gap. What would be a more pragmatic reading?
Would I really be gaining any noticeable difference with the "e"?
Of course I'm happy with my non-e but thinking if I could ramp up sharpness that much....

[Fine_Art]

Hi Arthur,

There are commercial Chrome-on-glass versions available for lens testing. However, they are not suitable for sampled / digital camera imaging, because the sharp edges add high spatial frequencies that will lead to aliasing artifacts. Also, the high contrast will not be needed, and can even distort the read-out (due to glare), and cause trouble in subsequent processing (more sensitive to built-in sharpening, causing overshoots and clipping).

These black/white targets are not suited for digital imaging (they are for analog imaging, vibration testing, or inspection of the lens on an optical bench). That's why I invented the sinusoidal version more than a decade ago (it was a 60 cycle version, because it was less demanding to print with the technology of those days).

Cheers,
Bart

I don't really understand that. I do believe you, maybe you can flesh it out a bit more. If we have a camera with a LPF, the aliasing should be minimal or software generated. If we have a camera like the 800E we should get sub-nyquist radius garbage which should give the user an idea of how the camera is going to work on fine patterns. If we look at a radius greater than the nyquist circle we can calculate the frequency relative to the pixel frequency. Of course this is basic to you, I am trying to explain what the idea was. So I am thinking that as the frequency scales out, you can also see how it attenuates from pure black/white. Isnt this useful? Even if the lens mangles it with glare, that is the system you have to deal with. Real objects dont have rounded edges, they just finish. I would expect that when you get to the point where the pixels are 1/2 the width of the bars you should be hitting pure black, pure white. At frequencies inside that circle there should be the whole MTF curve. What is wrong? This is not my field of study.

[Ray]

Now the tests are all shot wide open correct?
I never shoot wide open, if ever above f4 mostly 5.6-11.
Would any diffraction still equalize the bodies at smaller apertures.
Before I bought the d800 everyone said "the difference is so minor" "sharpening will minimize any difference"
These tests seem to indicate a pretty huge gap. What would be a more pragmatic reading?
Would I really be gaining any noticeable difference with the "e"?
Of course I'm happy with my non-e but thinking if I could ramp up sharpness that much....

I think it's already been mentioned that these significant, perceived differences in resolution and sharpness, between the D800 and D800E, apply to the unsharpened images that have resulted from using a particular lens at its sharpest aperture, whatever aperture that may be.

When one begins to sharpen images for comparison purposes, then one introduces a whole new ball game. There are so many parameters and variables in sharpening routines (just look at the sharpening options in ACR, and in Smart Sharpen in Photoshop), and also different sharpening programs, such as Focus Magic, each with their own strengths and weaknesses, that it becomes very difficult to achieve a truly objective and unbiased result when one introduces sharpening.

In other words, one type of sharpening that produces the best result with a D800 image, might not be the best type of sharpening to get the best result from a D800E image. And even if you've done wide-ranging experimentation to demonstrate that the resolution differences after sharpening are in the order of only 1%, someone else might come up with a different sharpening routine, either now or in the future, which might significantly widen that gap of 1%.

DXO understand this perfectly, which is why they deliberately avoid introducing sharpening techniques into their methodology.

Best wishes for a speedy resolution of Edmund's current medical problems.

[Bart_van_der_Wolf]

I don't really understand that. I do believe you, maybe you can flesh it out a bit more. If we have a camera with a LPF, the aliasing should be minimal or software generated. If we have a camera like the 800E we should get sub-nyquist radius garbage which should give the user an idea of how the camera is going to work on fine patterns.

Hi Arthur,
The problem is that also an OLPF equipped sensor will exhibit aliasing artifacts, only less prominent. The reason is that to eliminate ALL aliasing potential, it would require a much stronger pre-blur. It would result in a very low contrast and low resolution image. That's partly due to the different sampling densities between the green and the Red/Blue channels of the Bayer CFA.

If we look at a radius greater than the nyquist circle we can calculate the frequency relative to the pixel frequency. Of course this is basic to you, I am trying to explain what the idea was. So I am thinking that as the frequency scales out, you can also see how it attenuates from pure black/white. Isnt this useful?

The problem is that the images will be hardly useful due to aliasing and related False color Demosaicing artifacts. See the attached simulation of a sinusoidal grating and a bi-tonal grating of the same star, side by side. First the regular images, then the same after overlaying a Bayer CFA filter in Photoshop and demosaicing the result in PixInsight software with a bi-linear demosaicing algorithm, then with a VNG algorithm instead of bi-linear. This would be similar to how a sensor array without OLPF would respond. The fourth attachment was Gaussian Blurred (0.3 radius) before overlaying with a Bayer CFA, and then demosaiced with the VNG algorithm just like the third attachment.

Real objects don't have rounded edges, they just finish.

Only in an 'ideal' world. However, with discrete sampling sensors, we are still faced with sensels that use area sampling, which averages the signal over the sensel aperture (IOW the edge will gradually go from covered to not covered, and the average is encoded as a square of uniform density). This in contrast to a point sampling system. Even our eyes, due to lens and inner-eye (vitreous, or gelatinous mass) imperfections blur the image forming rays to something more (Co)sinusoidal. At least our eyes use dithered sampling ...

I would expect that when you get to the point where the pixels are 1/2 the width of the bars you should be hitting pure black, pure white. At frequencies inside that circle there should be the whole MTF curve. What is wrong? This is not my field of study.

Just draw a circle of 92 pixels diameter (=Nyquist), and a circle of 183 pixels diameter (=0.5x Nyquist), on the attached renderings of the star centers for the relevant spatial frequencies.

Cheers,
Bart

[eronald]

Best wishes for a speedy resolution of Edmund's current medical problems.

Thanks. I'm ok, but for a change feeling about as smart as I really am
Going to watch this thread scroll by for another day, before I become my usual bitchy self.

Edmund

[Fine_Art]

Bart,

Thanks for the clear explanation, that really fills in some gaps.

[Alan Klein]

I'm a layman and don't understand much of what's been said above.  I mainly shoot landscapes.  If I was to buy a new camera, would I be better off with the D800 or D800E and why?

[Jim Kasson]

Thanks. I'm ok...

Glad to hear it.

Going to watch this thread scroll by for another day, before I become my usual bitchy self.

Can't wait.

Jim

[Jim Kasson]

IIRC, others (e.g. Frans van den Bergh with his MTF Mapper application) have tried to model an AA-filter with four (Gaussian or Box?) blur kernels, and a variable (orthogonal? +) diagonal offset between these kernels.

Thanks for the pointer, Bart. Jack Hogan also gave me a link to the same site. I threw the 4-way beam splitter into the model, and found that it did much better than a straight box blur (modeled by making the fill factor 400%) in the stop-band, and a little worse in the pass-band.

Details here: http://blog.kasson.com/?p=5832

Jim

[Jim Kasson]

I'm a layman and don't understand much of what's been said above.  I mainly shoot landscapes.  If I was to buy a new camera, would I be better off with the D800 or D800E and why?

The first test is whether you're willing to go to all the trouble to get the most out of the E version: sturdy tripod, good lenses, mirror lockup, live-view focusing, keep the focal length of the lenses below, say, 300 mm... You get the idea. If you're OK with that, the next test is whether you want your files to be mostly aliasing-free with default settings in your raw converter of choice, or whether you're willing to do some work on some of them to clean up the false color. If you are willing to do some editing and you said yes to the first set of conditions, the E is a good choice. If you want your files to be mostly free of moire and false color just as they came from Lr, then you probably don't want the E.

However, as Bart pointed out above, there's a condition where the E is a better choice than the regular D800 that's counterintuitive: if you're going to have images where the combination of lens defects, diffraction, focus errors, vibration, etc are enough to keep you from having aliasing, you're better off without the extra blur of the AA filter, so you want the E.

Confusing, huh?

Jim

[Jim Kasson]

I have always wondered why digital imaging doesn't use a figure like total harmonic distortion (THD) like in audio.

THD is a measure of departure from linearity. So far, in this discussion, we've assumed linearity, so I'm not sure what your point is here. Maybe you could expand on that.

In my experience, if you stay away from electron counts near the full well capacity, where there is oftem some compression, today's sensors are remarkably linear.

Even if you had a perfect brick-wall low pass filter you will not get a high quality signal at or very close to Nyquist frequency

I'm assuming that your filter comment is unrelated to your THD comment, so I'll assume linearity here. I don't know what you mean by a high-quality signal in this context. Certainly, one of the consequences of steep filter skirts is large phase shifts. That's important in audio. Is that what you mean?

I've wondered about the effect of spatial phase shifts in imaging before, but have not reached any conclusions. They are explicitly ignored in MTF analysis. Is that a bad thing?

Jim

[Bart_van_der_Wolf]

I've wondered about the effect of spatial phase shifts in imaging before, but have not reached any conclusions. They are explicitly ignored in MTF analysis. Is that a bad thing?

Hi Jim,

Good or bad, maybe ... In the ISO SFR analysis based on a Slanted edge, the method averages into one quarter sensel bins, and enforces an 'exact' multiple of phase rotations. That tends to produce a stable result under all circumstances, but it hides e.g. a stair-stepping tendency of aliasing.

With the tests I've done on my Slanted edge evaluation method, I take fewer phase rotations (1/10th up to how many one has the patience for, I recommend 10/10ths), because it reveals such intricacies. As can be seen from the attached example, it takes more work, but it starts to show that for the sharpest aperture there is a slightly higher phase rotation effect than for the more blurry apertures. My results only fluctuate a bit, also because I mostly ignore Gamma, but the differences are usually buried in the fractional digits.

Cheers,
Bart

[Jim Kasson]

I've added a chart as attachment to show the trade-offs between lens quality (expressed as least achievable blur), sensel pitch, and resulting resolution in (simulated) Cycles/mm at MTF50. Diffraction is not fully modeled in, so that may reduce the amplitude at narrower sensel pitches a bit.

And here's a surface that looks a lot like Bart's for a diffraction-limited (red lambda 650nm, green lambda 550nm, blue lambda 450nm) lens (well modeled except for the discrete wavelengths: the Airy filter kernel was 3 times the diameter of the first zero times the target to sensor ratio of 32, and the run to generate this took about 4 hours) at from f/2.8 to f/17.5, and sensel pitches from 2 to 6 um. A RGGB sensor with Adobe RGB primaries was assumed, and bi-linear interpolation was used for demosaicing.

Z axis is MTF50 cycles per picture height, assuming a 24x36mm sensor, computed using sfrmat3.



Jim

[Fine_Art]

The first test is whether you're willing to go to all the trouble to get the most out of the E version: sturdy tripod, good lenses, mirror lockup, live-view focusing, keep the focal length of the lenses below, say, 300 mm... You get the idea. If you're OK with that, the next test is whether you want your files to be mostly aliasing-free with default settings in your raw converter of choice, or whether you're willing to do some work on some of them to clean up the false color. If you are willing to do some editing and you said yes to the first set of conditions, the E is a good choice. If you want your files to be mostly free of moire and false color just as they came from Lr, then you probably don't want the E.

However, as Bart pointed out above, there's a condition where the E is a better choice than the regular D800 that's counterintuitive: if you're going to have images where the combination of lens defects, diffraction, focus errors, vibration, etc are enough to keep you from having aliasing, you're better off without the extra blur of the AA filter, so you want the E.

Confusing, huh?

Jim

Why cant they make a screw on filter for when you need the AA? Just like putting a UV filter on the lens.

[Jim Kasson]

Why cant they make a screw on filter for when you need the AA? Just like putting a UV filter on the lens.

You could try Vaseline...

[Fine_Art]

You could try Vaseline...

Or panty hose.