24-90mm Focus Shift (Diglloyd)

[tashley]

Advertisement (gone after registration)

24mp of mushiness is less resolving than 19mp of pristine quality upscaled to 24mp.  So I wouldn't equate mp count with resolution.  

 

You are absolutely correct.  However I wager that at least some members here will be surprised by the quantum. We can disagree as to how much it does or doesn't matter but more information is good, right?

[Advertisement]

The Leica Q3 is available for pre-order at B&H Photo Video and Adorama

[cpclee]

Agreed

[Jared]

OK, here goes - and it seems worse than I suspected.

 

The first file is a JPEG export at full size of the DNG as shown in Lightroom having been imported with a sharpening profile I generally use but with no changes to image dimensions. It is 6000 x 4000 pixels (i.e. 24mp) -  as you would expect. Lightroom has clearly applied the lens corrections proposed to it by Leica and there is no option to un-apply them.

 

The second file is from a TIFF export from RAwDigger, showing the file with no lens corrections applied whatsoever. It is, as far as I can tell, RAW RAW. It measures 6016 x 4016 - which is 24.16mp.

 

The third file is processed from the same source as the second file so as to match, as closely as possible, the first file. The involves distortion corrections and then cropping. As you will see it is fairly close but not identical to the DNG. I can't get it exact because the lens correction profile provided by Leica is more complex than that available using manual controls in LR but I hope you will agree that it is at least a reasonable approximation. It measures 5314 x 3548 pixels which is equivalent to a touch under 18.9 mp.

 

 

 

No, I think you are interpreting it correctly.  I assume your samples were at the 24mm end?  The difference between the 6016 x 4016 and 6000 x 4000 is almost certainly effective vs. actual pixels.  That part doesn't bother me at all.  The 24-90 lens clearly has a ton of barrel distortion at the 24mm end that needs to be corrected in software.  This necessarily requires interpolation.  Effectively, you are giving up resolution in the corners and edges when the software makes these kinds of adjustments.  Center sharpness is likely not affected at all, but you will necessarily have a loss of detail elsewhere.  Is it enough to matter?  Probably not for most images, but I can certainly think of subjects where I would care.  That being said, I don't care about the megapixel count or how you would need to re-interpret the megapixel count at various focal lengths with various lenses.  Frankly, that wouldn't be very meaningful since it would differ so much from lens to lens.  In other words, I wouldn't consider it a "cheat"--from Leica or anyone else that they still list it as a 24mp camera since, with other lenses, it is.  A better way to account for this if Leica is being fair is to make sure the MTF data include the loss of resolution you get from the distortion correction.  No way to know whether they did for sure, but I suspect that they didn't--that the MTF is lens alone (with all the distortion still in place).

 

Well, I guess now we know why the technical brochure for this lens doesn't include distortion information.  Just MTF data.  Frankly, I was expecting that with a lens so large and heavy that Leica had done virtually everything in the lens, not in software.  Clearly that's not the case.  Thanks for doing the leg work on this!

 

Also, it looks like you have found part of the explanation for an earlier post in which someone encountered physical vignetting with their lens while taking pictures of, I think, a grand child in the snow.  Looks like for some reason the raw software failed to apply corrections, and the mechanical vignetting from the shade or end of the barrel was visible in a few of his images.  I guess we can stop claiming his issue was imagined.

 

- Jared

[tashley]

Jared, you voice many of my thoughts and questions? Does it often matter? No. Did I think that the point of such a large and design was to minimize these sorts of compromises? I guess I sort of assumed so. Most importantly, what's the factual background of the MTF? 

 

Ps yes, this was ar 24mm and f5.6

[scott kirkpatrick]

There was a long discussion of the size and nature of the distortion corrections back in late October 2015 when the first DNG files were available to inspect.  Horrible pictures, but you can see the gory details of the actual corrections contained in the DNG files at http://chromasoft.blogspot.co.il/2015/10/how-much-lens-correction-is-there-on.html  (also read the comments) for a 24mm focal length example.  I spent some time using Capture One (which makes the corrections optional, unlike LR) seeing how much more image is available if you leave the edges uncorrected.  I was able to save more than Tim did in his example.  But I can't find that October thread here at the LUF at the moment. 

 

You can find the distortion value at the end point (0.5,0.5) for the 24-90 worked out in my comment on Sandy's blog.  This is before applying distortion correction.  After correcting this value is taken to zero distortion at each focal length.  The expression that I evaluated is the curve that would have been plotted for the distortion curve as a function of radius with an optical only lens.

 

scott

[tashley]

Interesting Scott, thanks! I saved as much as I could in order to match the DNG, not to save as much as I could - on the basis that Leica would be saving as much as it could - but clearly the simple linear distortion correction in LR can't go as far as a correction profile which is bespoke to the lens. IMHO even with the wind in the fairest direction the file is unlikely to be more than 20mp at most and I suspect not even that. Take a look at the files I linked above if you have a moment and see what you think, I;d be very interested in a second opinion!

[scott kirkpatrick]

Advertisement (gone after registration)

If I can remember Sandy's explanation, what he called the "Adobe prescription" is:

 

First, invert the distortion curve so that you solve for the sources in the object plane where the light comes from that has fallen into each of the 24 MPx of the image that you captured.  Then spread those 24 MPx into an extended virtual array with the pixel spacing of the actual image chip, but extending out beyond it.  The extended virtual array will render lines as nice straight lines, and will probably have more pixels than you started with.  Finally truncate this virtual plane to 24 MPx, and that is the corrected image. 

 

Why 24 MPx?  Any other choice will cause expensive questions to customer support and quibbling on forums.

 

Make sense?  I found with the parameters that Leica provided, you can make the corrected rectangular image larger than 24 MPx if you like.

 

scott

[tashley]

 

 

Make sense?  I found with the parameters that Leica provided, you can make the corrected rectangular image larger than 24 MPx if you like.

 

scott

 

Sure, one could make it any size merely by up-resing. The question is, if you correct the distortion in the raw RAW file to the same degree as Leica's profile tells Lightroom to achieve and then crop the image so no white borders or black corners are showing, what pixel dimensions do you get without then doing the upresing step? I still think it's about 19mp. 

[scott kirkpatrick]

I suspect RawDigger gave you a file that has been down-ressed in order to fit the ugly corners into approximately 24 MPx.  If you would point me at the dng that you started with, I'll see what Capture One does with it.

 

scott

[tashley]

I'm not sure how much that will tell you Scott. I don't know how familiar you are with RAWDigger but it gives you the 'raw' raw file and lets you analyse it. It does nothing else other than allow a TIFF export if you have the pro version. It's the purest way I know of looking at what actually comes off a sensor though of course it can't undo pre-cooking in camera.

 

C1 (with which I'm pretty familiar) 'does more' to the file  - for example even with all corrections turned off it still shows the pixel dimensions as 6000 x 4000. In other words even if you turn off everything you can turn off, it is still 'developing' the file.

 

Raw digger isn't for 'developing' a file, it makes no attempt at that. It's for digging around under the hood to see what actually is in the file before any collaborations with RAW development programs like C1 or LR apply any adjustments, which one rarely has any control over. As such I really do think the truth is that the pure raw file is 6016 x 4016 and the other numbers are as in my original post above.

 

Here's a link to the file.

 

http://tashley1.zenfolio.com/img/g430059511-o374332360.dat?dl=2&tk=QNGIY2nk7Uv1HSXkuhkS0V6KX09BHWD_z9AE7F9c9pE=

[tashley]

I have just been looking at DPReview which has some raw files available for download from the A7RII and Gmaster 24-70 F2.8 and I've poked around under the hood of a RAW file and compared it to the JPEG. It seems that the correction in that case is losing about 0.41% at 24mm and F5.6 compared to my estimate of 26% for the Leica.

 

HOWEVER that might well be due to some pre-cooking of the RAW in camera. Which, if it is the case, is more opaque still than the Leica way of doing it.

[thighslapper]

I've tried this as well (SL 24-90 files)..... with Oloneo Photoengine (primarily an HDR programme).... that also yields an unprocessed TIFF with no DNG corrections applied..... same size. 

 

Situation is identical to that on the Xvario and T ...... some very crafty image correction ...... including eliminating moire and purple edging etc etc.

 

Despite a lot of tinkering I cannot get my image to be anywhere close to the Leica applied DNG corrections ...... it is miles better ....... and there seems to be no loss of detail or added distortion/artefacts even in the peripheries.

 

Whatever Leica are up to there seems to be no downside to any of this manipulation. 

 

Much as I would love the 24-90 to be optically perfect at all settings and 2.8 throughout ..... I bow to Leica's century of experience in lens production ........ if they could have done this without major problems I am sure they would have.

 

Leica's back catalogue is sprinkled with very fine lenses that lost money or were so complex that assembly was a nightmare...... they are not stupid 

[CheshireCat]

Modern lenses are designed specifically for use with their matching brand bodies and that's all that you can expect.

 

 

I use all my Canon lenses on a Sony A7R2, even if all of them were designed for their matching Canon bodies. They all shine one the 42 MP sensor, and the AF is working fine. I even get correct EXIF data.

Luckily, Canon engineers are not taking shortcuts (yet).

[CheshireCat]

kind o

The first file is a JPEG export at full size of the DNG as shown in Lightroom having been imported with a sharpening profile I generally use but with no changes to image dimensions. It is 6000 x 4000 pixels (i.e. 24mp) -  as you would expect. Lightroom has clearly applied the lens corrections proposed to it by Leica and there is no option to un-apply them.

 

 

 

Have you manually applied chromatic aberration corrections in Lr ? The raw file shows both lateral and longitudinal CA.

[scott kirkpatrick]

I look at the question of undoing distortion a little more flexibly than just using the Adobe protocol:

 

I read your dng into C1 release 9.  I goto the lens profile dialog, and turn on the crop tool.  I set "hide distorted areas" OFF, leave distortion corrected 100%, and now can see the same full frame that your RAWDigger output showed, with darkened corners and soft areas, etc.  The area filled is slightly pin-cushion shaped, since a lens with barrel distortion draws a pincushion into the rectangular frame.  The default "Adobe" crop gives you a wide band of image that is left out on all sides.  If I move the crop lines out so that they touch the edges of the image on all four sides, and render at 100%, I get a 4156x6433 jpeg that you can download from https://drive.google.com/file/d/0B2p1VlRvPAUBVTdHU1E1Q0N3WkU/view?usp=sharing .  I made no changes in C1 except for its default sharpening, and this restoration to a virtual grid of 26.74M 6 micron pixels, which would have been required to capture the rectangular image that is actually seen.  Actually, it looks pretty good.  It's not up-ressed, since the spacing in the center is the same as in the original image, but the image is extrapolated at the edges.  These 26.74 virtual pixels contain almost all of the information in the original nominal 24 MP, while the crop to 24 MP after removing distortion leaves only (24/26.7) or 90% of the original information.

 

One option that I use frequently with C1 is just to not remove the distortion, if there are no straight lines near the edges but only, say foliage.  That gives you all 24M lovely pixels, and requires no extrapolating.

 

scott

[IkarusJohn]

I see angels dancing on the heads of pins, I'd have to say.

 

Sure, AF can be hit and miss.  That's always been the case; and manual adjustment or fine tuning will always be more reliable.  But shock, horror the resolution at the corners when zooming out to 24mm is "only" 19MP rather than 24MP?  Really?

 

As I understand it, if that is true, 24MP -> 19MP is a reduction of 5MP in total, or a reduction of roughly 21% along the horizontal axis (a 19MP sensor being 4,750 x 3,170).  If one translates that into print size, if my maths is correct, the 24MP image (at 6,000 x 4,000) printed at 300PPI yields a maximum print size of 20" x 13.3"; whereas if the reduction to 19MP is correct, the maximum print size at the same print quality is 15.83" x 10.57"

 

I can appreciate that this might bother some, but to be honest it doesn't keep me awake at night.  If I'm wanting that size of print of that quality, I'd be likely to use either the 28 or 21 Summilux, a tripod, bracket for exposure and use focus magnification etc.  I'm not belittling these discoveries, just pointing out that we're really looking at the margins here.  For me, anyway. 

[tashley]

kind o

 

 

Have you manually applied chromatic aberration corrections in Lr ? The raw file shows both lateral and longitudinal CA.

 

Nope, that wasn't the specific point of this exercise though it would have been of additional interest.

[tashley]

I look at the question of undoing distortion a little more flexibly than just using the Adobe protocol:

 

I read your dng into C1 release 9.  I goto the lens profile dialog, and turn on the crop tool.  I set "hide distorted areas" OFF, leave distortion corrected 100%, and now can see the same full frame that your RAWDigger output showed, with darkened corners and soft areas, etc.  The area filled is slightly pin-cushion shaped, since a lens with barrel distortion draws a pincushion into the rectangular frame.  The default "Adobe" crop gives you a wide band of image that is left out on all sides.  If I move the crop lines out so that they touch the edges of the image on all four sides, and render at 100%, I get a 4156x6433 jpeg that you can download from https://drive.google.com/file/d/0B2p1VlRvPAUBVTdHU1E1Q0N3WkU/view?usp=sharing .  I made no changes in C1 except for its default sharpening, and this restoration to a virtual grid of 26.74M 6 micron pixels, which would have been required to capture the rectangular image that is actually seen.  Actually, it looks pretty good.  It's not up-ressed, since the spacing in the center is the same as in the original image, but the image is extrapolated at the edges.  These 26.74 virtual pixels contain almost all of the information in the original nominal 24 MP, while the crop to 24 MP after removing distortion leaves only (24/26.7) or 90% of the original information.

 

One option that I use frequently with C1 is just to not remove the distortion, if there are no straight lines near the edges but only, say foliage.  That gives you all 24M lovely pixels, and requires no extrapolating.

 

scott

Scott, with apologies for appearing contradictory, none of that works (if I am correct in my thinking). 

 

Firstly, C1 is a RAW developer and is not showing you the actual sensor data, however much you 'turn off'. You really do need Raw Digger to do that.

 

Secondly, extrapolation is interpolation is up-resing by another name: it is making information where there was none, but differentially across the image field. And crudely, too, since with corrections set to off, any complex field information that Leica is including in the DNG is not being used. I suspect that Leica's corrections use a reverse transformation which is far more refined than a 'one size fits all' distortion correction. My crop was not a default 'adobe protocol' crop, but it was a crop to maximise the pixel area of the crudely distortion corrected image whilst replicating the FOV shown in the Leica DNG in LR (and in OOC JPEG) whilst retaining the original aspect ration. There's no other sensible (less crude) way to do it because the only way to achieve the subtle, Leica-provided correction (transformation) is also to accept the crop and the up-res with which it comes packaged.

 

Thirdly, though I might be misunderstanding you, you seem to be effectively up-resing (to 26.74 mp - that HAS to be an up-res since it has more pixels than the sensor size) then distortion correcting and then cropping to get 24mp. it seems to me that I could up-res it to 100mp then distortion correct it and get a 90mp file. The math is relative to the fact that it has been up-resed.

 

So AFAIK the only way to do this accurately is to import into Raw Digger, export that 'raw' RAW as a TIFF then distortion correct it to match the DNG as closely as possible, then crop to the same subject matter as the DNG as seen in LR. Very important that: the DNG in LR gives you pretty much what an OOC JPEG gives you and what Leica intends you to see*. Measure the pixel dimensions of that, and you have the mp basis from which Leica instructs its JPEG engine and Lightroom (bundled with the camera)  to deliver via an up-res, which is 24mp.

 

* your crop, provided in the download you linked to, is not the same (yours is much wider) as the DNG as seen in LR and therefore also much wider than OOC JPEG - therefore it does of course have a larger mp count but it also has blackening corners at bottom and weaker than 'Leica would like you to see it' sharpness at the edges... and it has a different aspect ratio...  in other words you have chosen to include bits that Leica sees fit to throw away.

 

My aim in this exercise wasn't to squeeze the last drop but to replicate the Leica-approved version and thus to see how many real megapixels it is resed up from, and I still think that is <19mp  !

[Advertisement]

Hallo,

Look here: Leica L-Mount

[Jared]

Tim, thebonlynpart Inwould quibble with is the characterization of the image as a "true" (my quotes) 19 megapixel file up-resed to 24 megapixels. While you are correct that distortion correction inherently involves interpolation, the loss of resolution will not be distributed evenly. Corners and edges will be degraded, but the middle of the frame will still contain all the same information that a "true" 24 megapixel image would have. I doubt one could measure a meaningful reduction in resolution across the center 2/3rds or so of the image.

 

I would love to see MTF data both with and without the distortion correction. I don't think it is reasonable to think of this in terms of a megapixel equivalent since the files are still 24 megapixels in size, and almost all cameras are heavily interpolated anyway because of the Bayer matrix. The problem, of course, is that we consumers don't have the tools at hand to generate MTF data which is, I suppose, why you have resorted to a megapixel equivalent.

 

I guess the specific numbers don't really matter, anyway. The real question is whether the camera/lens combo delivers the results you want and expect at the 24mm end. In my case it does. I like the pictures and find that they have sufficient detail where I need it as long as my technique is solid.

 

- Jared

[scott kirkpatrick]

There's really not a lot of mystery to "Leica's complex field information," contained in the DNG.  The specification for interpreting the manufacturer's distortion and LCA correction parameters is given in Adobe's documentation for the DNG format.  There is a free download of the DNG SDK tools that lets you extract the parameters.  Sandy puts down the "gory details" in the ChromaSoft blog post that I linked to yesterday.  There are four parameters for each color plane, labelled a0, a2, a4, and a6.  The distortion curves that we used to see in lens technical reports as a function of distance from the center of the image are evaluated with these to give the radius r' from which the pixel that we see at r in the distorted image came:

 

r' = a0 + a2*(r^2) + a4*(r^4) + a6*(r^6)

 

Differences between the values of the ai for the three colors seem to be very slight, so average them, evaluate this function, and you have the distortion curves that we used to see.  They are missing from the current Leica documents because Leica believes we should prefer to correct them to zero after demosaicing the image.  Notice that with a sixth order polynomial you can describe moustache distortion, the curves which give the greatest distortion at intermediate radii, decreasing as you go to the edges.  This is commonly found in classic "optical only" designs.

 

I haven't worked with RawDigger, but if I read your DNG into C1 and reduce the distortion correction from 100% to 0%, I see a 24 MPx demosaiced image that corresponds to what you see with RawDigger.  With distortion correction set to 100%, each point in the original image is placed at its correct location at a radius r' given by the polynomial (for its color), and then mapped into an extended grid of virtual pixels with the original spacing.  It takes 27 or so MPx to hold all that information, and now the edges of the virtual image are bowed out, but that is the information captured from the true image.  All that I am saying is that it looks worthwhile to use more than 24M of those virtual pixels if you want an image that has straight lines near the edges, but this will require some extrapolation (only near the edges).  In C1, I can avoid the extrapolation if the content of the image doesn't require correction or benefits from some distortion, and will get ever so slightly crisper grass, gravel or leaves at the edges.

 

scott

 

P.S.  do it yourself MTFs or at least curves of contrast versus spatial frequency are possible using Imatest and a target image of tilted black squares on a white background.  It's fun the first time, but after that it does feel like angels dancing...