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

[eronald]

The point of a low pass filter is to cut off at a certain max frequency. Instead, it seems the D800 filter just chops 25% or so off the measured resolution (whatever that means) of each lens.

http://www.dxomark.com/Reviews/Best-lenses-for-the-Nikon-D800E-The-sharpest-full-frame-camera-ever-measured/Sharpness-analysis-D800E-vs.-D800

Edmund

[BernardLanguillier]

I guess I'll add a D800S to my line up.

The breakthrough Nikon achieved 2 years ago continues to be revealed it seems.

Cheers,
Bernard

[Paul2660]

+1. But I do hope to see wireless on the D800s native.

Paul

[David Anderson]

I saw that, but does it contradict some of the other information around on the resolution differences between the two models ?
Anyway, it's an amazing camera the E..

A newer model with a bit of an update would be welcome here..

[Jim Kasson]

The point of a low pass filter is to cut off at a certain max frequency. Instead, it seems the D800 filter just chops 25% or so off the measured resolution (whatever that means) of each lens.

AFAIK, we don't have anything near a brick wall AA filter. I think the only way to get there is to oversample with no AA filter and res back down, adding the AA in with digital filtering.

However, to say that current AA filters are possibly a joke without talking about aliasing is to focus on the disadvantage (loss of energy at spatial frequencies below the Nyquist frequency) while ignoring the advantage (reduction of energy at spatial frequencies above the Nyquist frequency).

Take a look at this:

http://blog.kasson.com/?p=5809

If you want to suggest possible AA filter convolution kernels, I'll try them out in the camera simulator and report the results. If you want to use negative numbers you can get some interesting results, but I'm not sure how realistic that is.

Jim

[Dave Ellis]

The problem I have with the DXOMark Perceptual MP concept is that I have not seen a suitable explanation of exactly how it is calculated and as a result I am unable to appreciate the significance of the measurement. I prefer to talk in terms of mtf vs lppmm.

The purpose of the AA filter is to restrict spatial image energy below the Nyquist frequency of the sensor as much as possible to minimize the creation of sampling artefacts. I don't know how camera designers decide on the strength of the filter but presumably in a camera like the D800, it would be tailored to have the smallest possible impact on sharpness with the sharpest available lenses. (Lower quality lenses provide extra low pass filtering anyway). With the D800E, the only filtering by the camera results from the sensor fill factor and the de-mosaicing algorithm used. The issue with the AA removed is how much energy above the Nyquist frequency the lens lets through and hence the likelihood of artefacts. From what I've read, the D800E is pretty good in this regard. I guess it is not particularly critical with a lot of landscape scenes and also with these it is quite likely that a fairly small aperture is used resulting in a lower mtf for the lens anyway (and hence extra low pass filtering).

Dave

[eronald]

AFAIK, we don't have anything near a brick wall AA filter. I think the only way to get there is to oversample with no AA filter and res back down, adding the AA in with digital filtering.

However, to say that current AA filters are possibly a joke without talking about aliasing is to focus on the disadvantage (loss of energy at spatial frequencies below the Nyquist frequency) while ignoring the advantage (reduction of energy at spatial frequencies above the Nyquist frequency).

Take a look at this:

http://blog.kasson.com/?p=5809

If you want to suggest possible AA filter convolution kernels, I'll try them out in the camera simulator and report the results. If you want to use negative numbers you can get some interesting results, but I'm not sure how realistic that is.

Jim

Jim,

 Where are details on your simulator?
 
Edmund

[fdisilvestro]

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

[Jim Kasson]

Jim,

 Where are details on your simulator?
 
Edmund

Edmund, I've been describing its capabilities on my blog. It's written in Matlab. I can email you some code if you'd like; it's pretty disorganized now, as I'm working on it every day. I'm using Peter Burns' sfrmat3 to get the MTFs. WRT the AA filter simulation, all I need is a convolution kernel at at least the resolution of the camera's pixel pitch, or preferably some higher resolution. I'm still working on the right way to do AA filter simulation, and having some examples to try would be instructive for me, as well as motivating.

Jim

[eronald]

Edmund, I've been describing its capabilities on my blog. It's written in Matlab. I can email you some code if you'd like; it's pretty disorganized now, as I'm working on it every day. I'm using Peter Burns' sfrmat3 to get the MTFs. WRT the AA filter simulation, all I need is a convolution kernel at at least the resolution of the camera's pixel pitch, or preferably some higher resolution. I'm still working on the right way to do AA filter simulation, and having some examples to try would be instructive for me, as well as motivating.

Jim

What would be nice would be to have some small mini-simulator with actual blocks, either fourier or spatial, not just MTF. I proposed this to Iliah a few years ago, he basically laughed, but we have a great number of raw converters and zero public domain full system simulators AFAIK

Edmund

[Jim Kasson]

What would be nice would be to have some small mini-simulator with actual blocks, either fourier or spatial, not just MTF. I proposed this to Iliah a few years ago, he basically laughed, but we have a great number of raw converters and zero public domain full system simulators AFAIK

Edmund, I'm not sure what you mean by actual blocks; do you mean code modules?

Let me tell you how my simulators -- I've built two of them, one aimed at camera noise studies, and one that started out looking at mosaicing and demosaicing, and grew to handle the Sony lossy raw compression, diffraction, AA, lens defects and other things -- work. Actually, let me just concentrate on the second one. It starts with a target image, which can be anything as long as it's at lest a couple of binary orders of magnitude larger in each linear dimension than the simulated sensor. Then I simulate the lens defects through convolution. Then I simulate the AA filter the same way. Then I sample to a (just RGGB, for now) Bayer CFA with just about any binary fill factor, which can exceed 1, adding photon noise if desired. Then I add gain from the ISO knob and digitize with an arbitrary resolution ADC. I demosaic with bilinear interpolation, and can write out the resultant image, or, in the case of a slanted edge target, analyze it with sfrmat3.

There are some significant limitations at this point. The camera's CFA is that which yields the Adobe RGB primaries. Diffraction is computed for any wavelength, but is the same for all color planes. I can't write out raw images in a form that DCRAW can deal with. Only on-axis lens defects are considered, and they are crude.

The first simulator was object oriented, but the second one, which just grew in unplanned directions, is messily functional.

FWIW...

Jim

[hjulenissen]

The point of a low pass filter is to cut off at a certain max frequency. Instead, it seems the D800 filter just chops 25% or so off the measured resolution (whatever that means) of each lens.

http://www.dxomark.com/Reviews/Best-lenses-for-the-Nikon-D800E-The-sharpest-full-frame-camera-ever-measured/Sharpness-analysis-D800E-vs.-D800

Edmund

Most of us use our cameras for recording images that we watch. Measurements are only interesting as long as they are relevant.

Undocumented measurements of "sharpness", while interesting, can not tell us if a particular camera feature is "a joke".

(yes, I feel that both the dxo article and your post is somewhat populist)

I have implemented (near) brickwall filters for images, and they look horrible. Even the most popular filter kernels for digital image scaling (where you have more freedom than doing physical/optical filtering) does some harm to the pass-band in order to limit ringing etc. Nearest neighbor provides excellent "sharpness" but generally horrible image quality.

-h

[Ray]

The point of a low pass filter is to cut off at a certain max frequency. Instead, it seems the D800 filter just chops 25% or so off the measured resolution (whatever that means) of each lens.

http://www.dxomark.com/Reviews/Best-lenses-for-the-Nikon-D800E-The-sharpest-full-frame-camera-ever-measured/Sharpness-analysis-D800E-vs.-D800

Edmund

Wow! I'm sure glad I chose the D800E in preference to the D800, and ignored all those experts whingeing about the greater potential of the D800E to produce aliasing artefacts. 

[Bart_van_der_Wolf]

Edmund, I've been describing its capabilities on my blog. It's written in Matlab. I can email you some code if you'd like; it's pretty disorganized now, as I'm working on it every day. I'm using Peter Burns' sfrmat3 to get the MTFs. WRT the AA filter simulation, all I need is a convolution kernel at at least the resolution of the camera's pixel pitch, or preferably some higher resolution. I'm still working on the right way to do AA filter simulation, and having some examples to try would be instructive for me, as well as motivating.

Hi Jim,

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.

While an interesting exercise, I feel that the interactions between optical components are of such complexity, and so many variables are involved (e.g. irregularly shaped sensel aperture, hence complex shaped fill-factor), that an inaccurate model of one variable, may lead to false conclusions about the whole system (just like one cannot simply multiply the MTF of diffraction blur with the MTF of defocus).

In fact, the complexity of compound blur kernels is such that also Frans seems to use Monte-Carlo simulation to get a statistical approximation because it would presumably take too long to accurately calculate the convolution of all components.

In addition to that, I keep finding (empirically) that combined with Raw conversion algorithms the cumulative/compound blur of all involved components closely resembles a Gaussian blur kernel (or a combination of a few different Gaussian blur kernels for complex blur shapes). Others use the Moffat function (see also here) instead of a Gaussian, to allow a more accurately fitting model of the shape of the compound PSF.

So maybe just adding a simple Gaussian blur will get you close enough to a realistic simulation, without the risk of over-specifying the approximation model.

Cheers,
Bart

[Bart_van_der_Wolf]

The point of a low pass filter is to cut off at a certain max frequency. Instead, it seems the D800 filter just chops 25% or so off the measured resolution (whatever that means) of each lens.

Hi Edmund,

The cut-off is very gradual, and affects all spatial frequencies (not a brick wall cutoff, if it were even possible, because that would introduce e.g. ringing artifacts). Also, the DxOmark single figure metric is a combination of perceptual and physical factors thrown into one basket. It also doesn't tell you anything about the potential for deconvolution restoration, for which the MTF response near the limiting resolution is more important than its weight in the Mpix metric suggests.

When the D800E became available, I tested it's resolution potential compared to that of the D800, based on a few images that were made available to me by Michael Reichmann. The difference in absolute limiting resolution was approx. 1%, although the modulation near that limiting resolution was lower for the D800, making it more of a challenge to deconvolve and restore resolution (especially for low luminance contrast micro-detail), although with fewer aliasing artifacts getting in the way.

So the DxO Mpix metric is heavily influenced by their perceptual component, but without capture sharpening.

I'd not take it as anything more than an indication (as all single figure metrics attempting to describe a complex system).

Cheers,
Bart

[eronald]

Bart,

 As you know I'm very superficial and tend to go off fast

 What disturbs me with the DxO results is that rather than an impression of frequency clipping, there is a seeming quasi-proportionality of the Non-E/E numbers. I was expecting there would be a cutoff effect, where bad lenses are not very affected by the filter, but their numbers appear to show a loss of rez that can somehow be imagined somehow proportional to the lens max rez/quality/cost. So a blurry lens gets even blurrier.

 A lot of people have blurry (legacy) lenses; they probably think there is no point in getting the "E"; from this study it would appear that to the contrary they *should* get the E.

 Maybe somebody can explain to this sleepy idiot what the scales are and what is being described by their numbers. Every time I look at something from DxO I get confused. Such pretty diagrams, such a dumb reader. 

 As an aside I had a conversation with the Fuji engineers at a press presentation during the last PK; during the presentation their manager said "our sensor, no Moiré! ". So I filled my mouth with random words, walked up and said "Hey, you say no Moiré, but I think you *must* have some form of aliasing, what happens if you have a spatial frequency at twice Nyquist?" and the guy without batting an eyelid told the stupid journalist "such a frequency is blocked by the lens"!

 Also, if I were feeling philosophical, I would wonder exactly where the energy that is filtered goes? Does it heat the filter? Are the photons absorbed by the lattice or re-emitted in some way?

Edmund

Hi Edmund,

The cut-off is very gradual, and affects all spatial frequencies (not a brick wall cutoff, if it were even possible, because that would introduce e.g. ringing artifacts). Also, the DxOmark single figure metric is a combination of perceptual and physical factors thrown into one basket. It also doesn't tell you anything about the potential for deconvolution restoration, for which the MTF response near the limiting resolution is more important than its weight in the Mpix metric suggests.

When the D800E became available, I tested it's resolution potential compared to that of the D800, based on a few images that were made available to me by Michael Reichmann. The difference in absolute limiting resolution was approx. 1%, although the modulation near that limiting resolution was lower for the D800, making it more of a challenge to deconvolve and restore resolution (especially for low luminance contrast micro-detail), although with fewer aliasing artifacts getting in the way.

So the DxO Mpix metric is heavily influenced by their perceptual component, but without capture sharpening.

I'd not take it as anything more than an indication (as all single figure metrics attempting to describe a complex system).

Cheers,
Bart

[Bart_van_der_Wolf]

A lot of people have blurry (legacy) lenses; they probably think there is no point in getting the "E"; from this study it would appear that to the contrary they *should* get the E.

Hi Edmund,

Yes, that is a common mistake.

The combined system MTF response is roughly a combination of lens and sampling system MTFs. An increase in either, will boost total performance. Better/denser sampling (with minimized aliasing artifacts) will also allow better Capture (deconvolution) sharpening. The possibilities of resolution restoration through deconvolution of clean signal data are still underestimated.

Cheers,
Bart

[eronald]

Hi Edmund,

Yes, that is a common mistake.

The combined system MTF response is roughly a combination of lens and sampling system MTFs. An increase in either, will boost total performance. Better/denser sampling (with minimized aliasing artifacts) will also allow better Capture (deconvolution) sharpening. The possibilities of resolution restoration through deconvolution of clean signal data are still underestimated.

Cheers,
Bart

Bart, maybe this deserves a hard look / article. There is something counterintuitive here.Of course, moving from the spatial to the frequency domain and back is never obvious, and here we also have the fact that MTF numbers are used which again brings the shape of the MTF curve into play etc.

Edmund

[Jim Kasson]

Bart, maybe this deserves a hard look / article. There is something counterintuitive here.Of course, moving from the spatial to the frequency domain and back is never obvious, and here we also have the fact that MTF numbers are used which again brings the shape of the MTF curve into play etc.

This shows what Bart's talking about. It's really crude, and I'll work on getting a better plot with more points, log2 scale on the f-stop axis etc, but I thought I'd post it while the topic is still current:



What you're looking at is simulated MTF50 in cycles/picture height (for a 24mmx36mm sensor) vs sensel pitch in um coming towards you, and f-stop for a simulated Otus 55mm going from left to right. No AA filter.

Jim

[Bart_van_der_Wolf]

Bart, maybe this deserves a hard look / article. There is something counterintuitive here.Of course, moving from the spatial to the frequency domain and back is never obvious, and here we also have the fact that MTF numbers are used which again brings the shape of the MTF curve into play etc.

Hi Edmund,

In addition to Jim's example chart for a lens quality (MTF) that varies with aperture on one axis, and sampling density (sensel pitch) on the other, one can see that with better lens quality (here it's due to optimum balance between residual aberrations and diffraction) AND denser sampling, the combined result gets better.

One could replace the aperture differences on one axis by different lenses at their optimum aperture on that same axis. Then combined response would still increase with denser sampling for all lenses, with a peak at the combination of best lens and most dense sampling.

To put it in other words,  at a certain level of detail with an MTF of 50% for a lens, and and MTF of 50% for a sensor,  the combined system response will be 0.5 x 0.5 = 0.25 MTF response (MTF25). If at the same level of detail the lens could be improved to 100% MTF, and the sensor is still 50%, then the combined MTF is raised to 1.00 x 0.5 = 0.5 MTF (MTF50, and it will never get better than the lowest contributor which therefore sets the maximum achievable limit, unless sharpening is added).

Likewise, if at the same level of detail the lens remains at 50% MTF, and the sensor could be improved to 100%, then the combined MTF is 0.5 x 1.00 = 0.5 MTF (it will also never be better that the lowest contributor). With both components at 100%, their combined MTF would become 1.00 x 1.00 = 1.00 MTF response (MTF100).

Therefore, increasing the MTF response for either component will raise the combined response. Of course, a 100% MTF for either component is virtually impossible (except for the lowest spatial frequency or with sharpening). It may also be easier to improve the MTF of one component for a reasonable price than for the other component. We also face physical limitations that prohibit 100% MTF, like diffraction and available sampling density. So it becomes an optimization problem for both components with bounds, and the lowest contributing component keeps dictating the best achievable combined response.

Cheers,
Bart

P.S. 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.