5Ds vs 5DsR - the true resolution difference?

[hjulenissen]

For those pining away for 4K video (and I'm not one of them), this isn't the camera to feature it.

I agree. If 4k video is important to you, looking at the highest sensel count stills-focused camera is a strange thing to do. Why would you want the manufacturer to go to the expense of reading out and processing >50 million sensels, only to downsample to 2000x4000? I mean, in due time this will perhaps be a standard feature of cellphones costing next to nothing, but right now this sensel count probably adds to the cost (and detracts from the flexibility) of a given product.

The Sony A7(s), rumored Canon 5Dmk4 and others in that class are probably better targeted at delivering good quality video (including 4k).

-h

[hjulenissen]

Hi,

The AA-filter reduces MTF, so that is expected.

The Canon 5Ds has higher resolution (in lp/mm) than any full format or MFD camera, so it may be a good candidate to use without AA-filter, although theoretical studies indicate that pixel pitches around 2.5 microns may be needed for correct rendition with very high quality lenses at near optimal apertures.

It would be interesting to take the theoretical/optimal findings to a more practical level. I am sure that it would also cause some dissent.

I believe that applying some degree of AA filtering can only reduce the magnitude of the MTF.

1. What kind of deconvolution is needed to make a given D800 image match a D800E image of the same scene, given optimal known technique, algorithms and parameter settings? How is the noise vs sharpness trade-off at high frequencies (perfect sharpness match might mean objectionable levels of high-frequency noise/artifacts). I believe that Bart has contributed on optimal gaussian deconvolution parameters, but how do you factor in the SNR at those frequencies in a visually enlightning way?
2. How close can a given image be matched in e.g. Lightroom for a D800/D800E comparision (assuming that commercial software is not necessarily optimal)
3. At what (if any) smaller apertures does the AA filter reduce sharpness by some non-trivial amount while diffraction ensure trivial amounts of aliasing even without the AA filter (is there a case for removing the AA filter for those who regularly want deep DOF and good detail, but may still not like aliasing)?
4. For a given AA filter blur (e.g. 0.7x sensel pitch horizontal and vertical comb filter), how large camera vibrations would be needed to obtain similar*) blur? For 50MP and e.g. 200mm lens, I have a gut-feeling that you have to bolt your camera really well to the ground for this to matter much, even if lens and focusing and aperture is ideal.

*)Camera movements and comb filters cause different blurring, but say that the camera is kept at 1/4 of the exposure at 2x2 different positions, corresponding to the "taps" of the AA comb filter.

[ErikKaffehr]

Hi,

The problem is that we don't have very much of comparable equipment. There are some Nikon users who own both the D800/D800E.

- What I can see is that I have a lot aliasing related problems on the P45+, 6.8 Microns no AA filter.
- There is some aliasing on my Sony Alpha 99 with 6 micron pixels and AA filter.
- With my Sony Alpha 77 I don't see aliasing problems.

But, I would also say that I don't feel the AA-filter is intended to eliminate aliasing, just to reduce colour aliasing and demosaic errors. The best way to reduce aliasing is to increase sampling frequency, that is reduce pixel size. At one point the AA filter will not be needed.

I don't think 4 microns are small enough, as Canon would not offer the Canon 5Ds with and without AA-filtering if AA-filtering was not needed in certain cases.

Regarding camera shake, it seems to be little problem with EFCS (Electronic First Curtain Shutter).

Best regards
Erik

It would be interesting to take the theoretical/optimal findings to a more practical level. I am sure that it would also cause some dissent.

I believe that applying some degree of AA filtering can only reduce the magnitude of the MTF.

1. What kind of deconvolution is needed to make a given D800 image match a D800E image of the same scene, given optimal known technique, algorithms and parameter settings? How is the noise vs sharpness trade-off at high frequencies (perfect sharpness match might mean objectionable levels of high-frequency noise/artifacts). I believe that Bart has contributed on optimal gaussian deconvolution parameters, but how do you factor in the SNR at those frequencies in a visually enlightning way?
2. How close can a given image be matched in e.g. Lightroom for a D800/D800E comparision (assuming that commercial software is not necessarily optimal)
3. At what (if any) smaller apertures does the AA filter reduce sharpness by some non-trivial amount while diffraction ensure trivial amounts of aliasing even without the AA filter (is there a case for removing the AA filter for those who regularly want deep DOF and good detail, but may still not like aliasing)?
4. For a given AA filter blur (e.g. 0.7x sensel pitch horizontal and vertical comb filter), how large camera vibrations would be needed to obtain similar*) blur? For 50MP and e.g. 200mm lens, I have a gut-feeling that you have to bolt your camera really well to the ground for this to matter much, even if lens and focusing and aperture is ideal.

*)Camera movements and comb filters cause different blurring, but say that the camera is kept at 1/4 of the exposure at 2x2 different positions, corresponding to the "taps" of the AA comb filter.

[hjulenissen]

I don't think 4 microns are small enough, as Canon would not offer the Canon 5Ds with and without AA-filtering if AA-filtering was not needed in certain cases.

Nikon would seem to think differently, as they are offering their D810 (4.88 micron?) only sans AA filtering?

-h

[ErikKaffehr]

Yes,

That may be the case.

On the other hand, the samples on Imaging resource show aliasing that is not present on the D800, check attached image. D810 on the left, D800E center and D800 right.



Best regards
Erik

Nikon would seem to think differently, as they are offering their D810 (4.88 micron?) only sans AA filtering?

-h

[Phil Indeblanc]

With the DSLR I am shooting "around F16". I actually have not taken that Leica Macro el100 lens out in some time. After I did my tests I came to the conclusion and put a few things away. So I'd have to retest.
I can test it with the Canon 180L Macro also, but I think at the time I found the Leica to very close or sharper/(contrast perhaps?). 

[Phil Indeblanc]

From that image: the 2 on the right can be better focused. They look like a small crop of a much larger image.
There is going to be some situations that moire shows up.

Maybe you are saying that after the size of pixel is less than 3 there is ZERO chance of moire/In relation to lens ? I think it may drop further, but not zero.

Either way, the image is printed with line screen, so it picked up that pattern. Its not that the artifacts were created from thin air. Most offset printing is done between 70-200line screen(150 average for color).

Oddly enough, years back I read an article that at some point sensors technology would slowly go to a single pixel with very high resolution. I think this would eliminate the problem you show :-)

This type of artifact on flat surface is very easy to fix, and I would rather have this problem to fix once in a blue moon vs not getting the optimal sharpness I can to start with.

[ErikKaffehr]

Hi,

F/16 would eliminate aliasing on pretty large pixel pitches. Personally I shoot mostly f/8 on DSLRs and f/11 on the P45+.

This is P45+ at f/8, 6.8 micron pixels:


And this is P45+ at f/16:


An SLT 77 at f/8 with 3.8 micron pixels:


The same image as above, properly downscaled:


What this shows is that:

• f/16 on P45+ pretty much eliminates aliasing.
• A small pixel camera like 3.8 micron pixel avoids aliasing mostly at f/8
• The small pixel camera actually resolves more details
• The image downscaled from the small pixel camera has better detail

All images from this article: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/78-aliasing-and-supersampling-why-small-pixels-are-good

Best regards
Erik

With the DSLR I am shooting "around F16". I actually have not taken that Leica Macro el100 lens out in some time. After I did my tests I came to the conclusion and put a few things away. So I'd have to retest.
I can test it with the Canon 180L Macro also, but I think at the time I found the Leica to very close or sharper/(contrast perhaps?).  

[Phil Indeblanc]

The small pixel 77 camera resolves more details vs the P45+ is what you're saying.

How is it that the P45+ file is smaller than the 77?

Can you clarify your statement about the 77 showing more detail vs the P45 at F16?  Is F16 the optimal sharpness for that lens?
I don't know why, but you dollor bill images for the P45 look low contrast.

Downsampling has an advantage. I would not consider that apples to apples.
Just this small difference you have is enough to fill in small detail areas in an image.
The top one is the 77 downsampled. You can see clearly the difference in file processing results.
I took measurements off the area just below the D and the 77 file has a lower blackpoint.


Your conclusions at the end is something I can agree with, No, moire does not happen often. I have been shooting with NON OLP filter sensors for close to 10 years and I think it happened once or twice on fabric. So for me the AA should only be an OPTION, as I would likely never use it. I think Canon should have had the courage to have zero AA filter, not a just a weak one.

[ErikKaffehr]

Hi,

The crops are actual pixels. The Alpha 77 has smaller pixels then the P45+, so the image is larger. Keep in mind that this is a pixel size demo, and not an MFDB vs APS-C comparison. Both images are shot with lenses of same focal length, at the same distance. Described in the article I was referring to.

The optimal aperture on the Hasselblad lens is probably f/5.6 - f/8, the image below is at f/8:



While the one below is at f/16, what is shown that artefacts are almost eliminated but at the cost of loss sharpness:



If you check the small pixel image, it contains much better detail than any of the above images. It was also shot at f/8.



Contrast differences probably come from processing variations, but stopping down to f/16 reduces global contrast.

Also, keep in mind that no well made lens for small or medium formats ever performed best at f/16. Large format lenses is a different case, as they are never used at large apertures. Any decent small format lens performs best between f/4 and f/8. This applies rom simple Nikkors to Coastal Optics and Otuses.

Left column on this page page demonstrates the effect of diffraction at different apertures on 4.8 micron pixel sensor (36 MP on full frame)
http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures?start=1

The lesson from this is that smaller pixels actually give more and better detail without generating fake detail, and can be used near optimal apertures without risking moiré.

This image shows that colour artefacts can occur on rippled water (P45+, Distagon 40/4 at f/11):


Best regards
Erik

The small pixel 77 camera resolves more details vs the P45+ is what you're saying.

How is it that the P45+ file is smaller than the 77?

Can you clarify your statement about the 77 showing more detail vs the P45 at F16?  Is F16 the optimal sharpness for that lens?
I don't know why, but you dollor bill images for the P45 look low contrast.

Downsampling has an advantage. I would not consider that apples to apples.
Just this small difference you have is enough to fill in small detail areas in an image.
The top one is the 77 downsampled. You can see clearly the difference in file processing results.

[Bart_van_der_Wolf]

Maybe the science says otherwise, and some tests use particular lens selects that show issues of moire without AA filters, but with the select gear I have and use....Importantly with subjects I work with, I have to say that the sensors I use without an AA create a noticeably sharper, more 3D raw file vs those with an AA.

The difference is something I see, and I cannot see post sharpening be the better alternate, nor be equal to getting the original lines (or as good) in the image.
Maybe its the macro subjects, but the difference is clear in my tests(DSLR, full, crop, DSLR without AA, MFdb without AA).

Hi Phil,

I'm also involved (but not exclusively) in Photomacrography, and can tell from experience that most folks shoot with apertures that are too narrow for quality, just to get some more Depth of Field. One then stacks the effects of diffraction on top of anti-aliasing filters, which is often too much. In such a case, it can help to avoid the use of an OLPF, although in my experience it can then rear the ugly head of aliasing in e.g. specular highlights.

Having said this I have respect for a number of posters here, and i'd like to see how I can better understand this.
There are a couple users here that have "scientifically" tested a few things, and I'd like to put theirs and my findings to see whats going on....Before I no longer have the gear/chance to do so.  Since in macro work the pixel peeping is not only a self driven desire, as the work has to be perfect and demands it. Editing has to be clean with a solid chunky file.

I'm open to testing, if you are. I can think of Erik and Bart....  and there are a few others not coming to mind at the moment...but we can setup some parameters to work under/?

Not sure what it exactly is what you want to test, but maybe the attached (but not a yes/no OLPF comparison) image I made for a different thread on TS lenses can still help a bit. It also shows the incremental benefit of using dedicated lenses. The recent edition of a € 5 note has an incredible amount of detail of varying contrast, and the (recent updated model) EF 100mm f/2.8L Macro offers an incredible amount of detail on a camera with OLPF, but one needs to use a relatively wide aperture (I used f/4.5) to avoid diffraction. That also means that I (need to) do a lot of focus stacking on stationary subjects to regain DOF. The 100% zoom crop represents an approx. 6.5x6.5 mm area in the original subject, shot at approx. a 1:2.4 ratio. I can't show a larger piece due to copyright restrictions, and Photoshop probably refuses to open complete reproductions of banknotes anyway .

My earlier tests on the difference between a Nikon D800 versus a D800E showed only a small, approx. 1%, difference in limiting resolution. I did see many more aliasing issues on shots of my cruel/merciless Star resolution test target from the OLPF-less camera model.

Cheers,
Bart

[Bart_van_der_Wolf]

It would be interesting to take the theoretical/optimal findings to a more practical level. I am sure that it would also cause some dissent.

I believe that applying some degree of AA filtering can only reduce the magnitude of the MTF.

1. What kind of deconvolution is needed to make a given D800 image match a D800E image of the same scene, given optimal known technique, algorithms and parameter settings? How is the noise vs sharpness trade-off at high frequencies (perfect sharpness match might mean objectionable levels of high-frequency noise/artifacts). I believe that Bart has contributed on optimal gaussian deconvolution parameters, but how do you factor in the SNR at those frequencies in a visually enlightning way?

Hi,

Not exactly sure what you mean with the SNR remark, but one obviously needs to use a deconvolution method that favors detail over noise, for an effective increase of the SNR. The regularized Richardson Lucy algorithm is still a very good algotithm, although for low noise images the regularized Van Cittert algorithm can also do wonders. Specific versions of deconvolvers are useful for spatially variant deconvolution across the image, e.g where the corners need very different parameters compared to the center of the image. Deconvolution can also be only applied on masked regions with a certain level of minimum detail, and thus spare smooth gradients from amplified noise.

2. How close can a given image be matched in e.g. Lightroom for a D800/D800E comparision (assuming that commercial software is not necessarily optimal)

Depending on the gap to close, the visual impression of contrast roll off towards higher spatial frequencies can be reduced by deconvolution. So then in output, it will be extremely hard to see any difference, unless the images are magnified a lot. Then the OLP filtered images will has marginally less absolute resolution, but it also depends on subject contrast. The magnification will also increase the visibility of aliasing artifacts, so it's a complex issue to answer, TANSTAAFL.

3. At what (if any) smaller apertures does the AA filter reduce sharpness by some non-trivial amount while diffraction ensure trivial amounts of aliasing even without the AA filter (is there a case for removing the AA filter for those who regularly want deep DOF and good detail, but may still not like aliasing)?

The effects of diffraction are not abrupt, so they will gradually increase as one stops down from wide open. It then becomes the interaction of the Lens MTF (limited by residual aberrations and defocus) and Diffraction that determines the effect on modulation. This is called the Optical Transfer Function (OTF), and its formula (which requires integation) is given e.g. by David Jacobson's excellent FAQ/tutorial.

In my experience, the Diffraction effects start to just become visually apparent when the diffraction pattern diameter becomes larger than 1.5x the sensel pitch. Others use a more relaxed criterion of 2x sensel pitch, but then I'm maybe more sensitive to seeing the degradation. From there on the diffraction will increase with the use of narrower apertures, until it completely destroys any modulation, at a spatial frequency of Cycles/mm = 1/(wavelength x Aperture), where wavelength is expressed in the same units as the spatial frequency, i.e. millimetres (e.g. 0.000555 for green filtered sensels, as explained here).

Slightly before diffraction alone reduces modulation/resolution to zero, the modulation will already be so low that only the highest contrast subject features stand a chance of being resolved, assuming the use of a perfect lens an perfectly circular apertures. More modest subject contrast will already lose enough system modulation to be unresolvable at lower spatial frequencies.

4. For a given AA filter blur (e.g. 0.7x sensel pitch horizontal and vertical comb filter), how large camera vibrations would be needed to obtain similar*) blur? For 50MP and e.g. 200mm lens, I have a gut-feeling that you have to bolt your camera really well to the ground for this to matter much, even if lens and focusing and aperture is ideal.

Obviously camera handling/technique will play a role, but it's hard to quantify, because so many other factors also play a rol. Defocus is one of the major ones, and there is only a very narrow range with truly perfect focus, the rest is just 'acceptable' (from a COC point of view, which is output viewing conditions related).

Cheers,
Bart

[Bart_van_der_Wolf]

I think Canon should have had the courage to have zero AA filter, not a just a weak one.

Hi Phil,

It's not a matter of courage. Modern lens designs have, as part of their optical design, the presence of a filterstack (IR filter, OLPF+Quarterwave plate+OLPF, sensor coverglass) included in the optimization calculations. By completely leaving out these optical elements from the optical path, the lens will now start to exhibit lots of deterioration from new residual lens aberrations. So to avoid having to use different lenses for different cameras, they chose to employ a OLP canceling method, just like Nikon did when they brought out the D800/D800E combo.

Cheers,
Bart

[hjulenissen]

Not exactly sure what you mean with the SNR remark, but one obviously needs to use a deconvolution method that favors detail over noise, for an effective increase of the SNR. The regularized Richardson Lucy algorithm is still a very good algotithm, although for low noise images the regularized Van Cittert algorithm can also do wonders.

My point was that if the AA-less camera produces what a given user wants (for a given scene), then the AA-filtered camera will have to be processed by deconvolution/sharpening in order to approximate that image.

If the SNR in spatial frequencies of interest is very low, then any algorithm to lift out the signal by a large amount is going to have a hard time. Thus, the difference between the 5Ds vs the 5DsR is not simply a matter of the linear spatial frequency response (difference), but also the expected SNR for a given scene/exposure in the significantly affected frequencies. If you can have 20dB SNR, then I assume that differences in sharpness can be reduced a lot. If you have 3 dB SNR, it might be hard to increase signal without excessive artifacts.

-h

[torger]

While I do think it's a good idea with OLPF and that sharpening can compenaste for added blur there is indeed a challenge of introducing noise as the OLPF-filtered image needs higher levels of sharpening. I mostly use simple USM sharpening (boosted with thresholds tricks), so I guess you can improve it with the more advanced sharpening packages. However I don't find a mild grain be as problematic as various aliasing artifacts so for my taste the increased noise is not much of a problem.

In the past I was thinking that diffraction could be a good replacement for an OLPF in landscape photography when you shoot with small apertures. I don't think that as much any longer. Diffraction blur is a soft onset with a large radius which means it's quite ineffective at killing aliasing, ie you need high levels of diffraction before aliasing disappears, and then you can have issues with global contrast and bleed around high contrast details. OLPF only blurs to the neighbor pixel as far as I know and to a higher extent so it's quite effective at killing aliasing while not lowering global contrast or bleeding too much.

With my 50MP MFD system I've found f/16 to be a suitable tradeoff between diffraction blur and aliasing removal, which corresponds to f/11 on the 5DS R. You still have aliasing left but a lot less than with one stop larger aperture, and it's still easy to sharpen without too much noise or artifact increase.

[MarkL]

To put it in perspective, I get more moire with my 12MP X100 than I ever have with my D800E. The issue is so small Nikon did away with the AA filter on the D810 (and the D7100) and this camera has even more resolution.

[Bart_van_der_Wolf]

My point was that if the AA-less camera produces what a given user wants (for a given scene), then the AA-filtered camera will have to be processed by deconvolution/sharpening in order to approximate that image.

Hi,

I think there may be a bit of a misconception in that assumption. The fact is that all images require (deconvolution) Capture sharpening, it's just that aliasing prohibits us from doing what is required on non-OLP filtered images (because that would exacerbate the aliasing artifacts). All lenses, even the OTUS types, have some residual aberrations and some diffraction in the mix. The filter stack and sampling aperture will also introduce some blur.

If the SNR in spatial frequencies of interest is very low, then any algorithm to lift out the signal by a large amount is going to have a hard time.

Yes, but you may be surprised at what is possible, assuming the original was properly pre-filtered. I'm not saying that Bayer-CFA images are properly pre-filtered or bandwidth limited, they aren't. But they do offer a better chance of surviving when they are low-pass filtered to some degree (however little or insufficient).

Cheers,
Bart

[Bart_van_der_Wolf]

While I do think it's a good idea with OLPF and that sharpening can compenaste for added blur there is indeed a challenge of introducing noise as the OLPF-filtered image needs higher levels of sharpening. I mostly use simple USM sharpening (boosted with thresholds tricks), so I guess you can improve it with the more advanced sharpening packages. However I don't find a mild grain be as problematic as various aliasing artifacts so for my taste the increased noise is not much of a problem.

In the past I was thinking that diffraction could be a good replacement for an OLPF in landscape photography when you shoot with small apertures. I don't think that as much any longer. Diffraction blur is a soft onset with a large radius which means it's quite ineffective at killing aliasing, ie you need high levels of diffraction before aliasing disappears, and then you can have issues with global contrast and bleed around high contrast details. OLPF only blurs to the neighbor pixel as far as I know and to a higher extent so it's quite effective at killing aliasing while not lowering global contrast or bleeding too much.

Hi,

Yes, diffraction is a rather poor substitute for a dedicated OLPF. The only benefit is that it blurs Red more than Green, but that won't matter that much in a luminance dominated demosaicing scheme.

Cheers,
Bart

[nemophoto]

Erik, thanks for posting those moire images of the beer label. Not sure how exactly the 5Ds and R would compare, because of everyone stating the smaller pixel photosites means fewer problems. But the label issue would definately be a no-go for me as a fashion shooter. I still have to deal with it enough with my 1Dx. Interestingly, my old 1Ds3 had far fewer issues than my 1Dx. I've always thought Canon weakened the AA filter on the 1Dx to lightly boost the res. Conversely, I don't really see much of a res gain on the samples of the Nikon 810 and 800E versus the 800. To my eye, what you really see is a slight contrast difference. I slight haze in a way. Easily correctable in post. One thing I've often said about digital versus film -- it's almost too sharp. I do far more retouching now than I did with scanned film! Especially with beauty.

To me, the biggest difference will be with in-store posters I shoot for one client. They are about 8 feet tall. Based on the files they request, (about 1/4 size at 400 ppi) I have to res up the final file anywhere from about 140% to 180%. The native 5Ds file would yield a file that would essentially be 100% most of the time. Hmm, I think I may need to boost my RAM from 32GB to 64GB! If I guess correctly, that will be a 500GB 16-bit file right out the gate, without ANY layers.

[Phil Indeblanc]

Interesting stuff, and funny enough we have covered most of this a few times:-)

If I want 1 tool doing most all tricks, and I know from my experience that it isn't the best way to "skin the cat".
Anyway, I have read for a few years, and seen it myself, but was never sure if that was the reason for the way the colors and lumes are laid out, but CCD vs CMOS I thought had something to do with SNR for clearer separation contrast?
(I hope I don't steer you into that topic entirely)