Sunday, October 18, 2015

Ugly Reality

I do try to avoid deep-ending on technical details, but the Zeiss Otus thing is sufficiently irritating that I'm going to do some math. Well, arithmetic.

We see, altogether too often, blather about lenses capable of delivering the necessary resolution to satisfy the demands of the newest digital cameras. Let's do some arithmetic, and see what these demands actually are. Take a 50 megapixel full frame camera. Bayer demosaicing throws away something like half of the resolution to give you color, so you're looking at 25 megapixels of resolving power. Doing some arithmetic, we come up with about 4100 pixels on the short side, which works out to 164 pixels per millimeter, or 82 line-pairs per millimeter. 82 lp/mm.

To put that in perspective, the reference film used for amateur lens testing, back in the day, was Technical Pan, which is capable of delivering 320 lp/mm of resolution. Yep. 4x the linear resolution. 16x the 2D resoluion. Now that's a challenging environment. Your 50 megapixel camera? It is to laugh. The much vaunted Reference Camera, the Nikon D810, clocks in a little under 70 lp/mm.

Having done some amateur lens testing in the 1990s, I can attest that actually getting 82 lp/mm to the sensor plane is hard. In reality, most people can't or won't accomplish it. The 50 megapixel wundercamera, while not in any meaningful way a challenging lens test case, has all the resolution normal people can actually use.

In the 1990s, any decent lens was capable of pushing 82 lp/mm onto the film plane. It is only in these modern times where, apparently, lenses have gotten so bad that they struggle.

So what the hell are all these people talking about? There are sensible people talking about how good the Otus lens line is.

Well, there IS something there. Consider some test charts. Suppose you have some black and white lines spaced very close together, so as to give you that 82 lp/mm resolution at the sensor. A very good lens will render them as crisp black lines alternating with white lines. A lesser lens will give you dark grey and light grey lines. The worse the lens, the lower the contrast at the sensor plane, until finally the contrast drops away to unusable levels. In the land of 82 lp/mm, both a very good and a quite moderate lens will deliver plenty of detail, but the better lenses will deliver higher contrast fine detail than the lesser one.

This mattered in the days of film, since recovering that contrast was technically demanding. Now, of course, it's the "Sharpen" slider. That's literally what that slider does. It doesn't make anything actually sharper, it puts the contrast back into the fine detail.

The better lens will also exhibit fewer aberrations, lateral blah and chromatic whatsit. Again, these mattered in the days of film, and are largely correctable in post in this digital age.

Bugger. Well, what on earth can this Otus thing actually do that is not doable with some minor digital corrections with a $200 nifty fifty?

Well. I guess there's fine detail that starts out low contrast? This can get lost, regardless of the sensor/film, when a cheap lens loses the thread of the very weak signal it starts with. This is and always has been the hardest thing to manage in the resolution part of the world, and even in the unrarified air of 82 lp/mm it can be a problem. Frankly, though, that's just a theory, and I suspect you'd be hard-pressed to actually find a situation in which the Otus preserved visible detail that a midrange equivalent lens lost.

Worth noting is that this kind of fine detail is by definition extremely hard to see in the first place, and hence arguably not very important visually, and it's quite difficult to preserve all the way to the print. So, you are (theoretically) saving something that nobody can see and which you will likely lose somewhere later in the process. Handy, that.

Secondly, correcting things in post will tend to add noise and artifacts. I dare say that fine detail from the Otus will look a trifle cleaner than fine detail from a midrange lens together with a touch of extra sharpening.

More importantly, it will look subtly different. So, when you're pixel peeping to see what your new $4000 lens is giving you, you'll be able to rationalize it.

There are other things like extreme field flatness, which is terribly helpful when you mainly take pictures of sheets of paper.

12 comments:

  1. Kodak Technical Pan data is here: http://www.kodak.com/global/en/professional/support/techPubs/p255/p255.pdf

    It resolves 200 cycles/mm, but only when used with a high contrast developer. If you need continuous tones, it resolves a lot less, and even that takes ideal conditions (no film curl, perfect focus, etc...).

    Bayer demosaicing throws away less than half the resolution, more like 20-30% on natural subjects.

    By direct comparison, Kodak Technical Pan (or rather Adox, which is a modern copy) is similar to a D810, when printed.

    Now about lenses: excellent lenses resolve 40 cycles/mm with something like 60-40% contrast full open (have a look at the MTFs from Zeiss), but only at the centre. True: close the aperture at f/8-f/11 and all 50mm lenses are equivalent. But that does not mean that lens defects will not show in many cases.

    And this is the reason for new lenses: with a D810 / A7RII, one can print at 24"x36" size (about A1). At these size, lens defects are quite visible, especially if you cannot close the aperture to f/8-f/11 or if you use wide angle lenses. This is the reason for better lenses or MF cameras.

    The Otus is a particular design not only designed for sharpness, but also to minimise longitudinal chromatic aberration (read the technical paper from Zeiss), which creates coloured fringes in out of focus areas and is difficult to correct in post. If one only wants sharpness, the Otus is overkill.

    ReplyDelete
    Replies
    1. Well, let's use your numbers, then. Tech Pan, the reference film for testing film resolution is 200 lp/mm, and the much-vaunted D810 leaps all the way up to 87 lp/mm, which is still a bit of a jump.

      If you're careful about selecting how much contrast you insist on, you can get any numbers you like, of course (which is how you got the 200 cycles/mm on Tech Pan from a chart that clearly shows useful response up to at least 250 before the line simply stops).

      I have my numbers, you have yours. Obviously, they come from different selection criteria. But I think we basically agree on the major points and are, inexplicably, quibbling over minutiae?

      Delete
    2. What I am saying is:
      -your figures apparently confused line pairs and lines. I talked about cycles, which are line pairs.
      -for continuous tone photography, Technical Pan in a real camera will not give you anything more challenging than a 24 mpix camera and film era lenses were also designed that way and not for 300 megapixels.
      -24x36 lenses do not have the resolution you believe they have, they are usually poor already at 40 cycles/mm, which is half the figure you noted
      -the defects of lesser lenses, particular wide-angle lenses used at fast apertures, will be quite obvious on large print sizes.

      So: no, we do not agree on major points. I know from experience that if you want to print big (and I do) and need to work in conditions where a fast lens is necessary, the better 24x36 lenses will give visually different results. So indeed there is a market for these lenses aside forum measurbators. And there I am not talking numbers but the much simpler experiment to give a file to print out big and look at it.

      Where I agree that if all you want is small prints or you can use f/8-f/11 most of the time, the cheaper lenses are all you need.

      Delete
    3. I did not confuse lines and line pairs. I do know the difference, and I did not confuse them. I did a quick search for Tech Pan's specs, and came up with 320 lp/mm, which is a perfectly credible number at some contrast, looking at Kodak's chart.

      We shall have to agree to disagree here on the remaining points.

      Delete
    4. Look at the Kodak document I cited at page 8, the figure on the right-hand column "Modulation transfer curves". Technidol is the developer for continuous tones, and you have 50% response at 100 cycles/mm. The curve stops at 30% response and 175 cycles/mm, but 30% response is lower than what is useful for general photography. The curve does not go up to 320 cycles/mm.

      Page 4, left-and column of the same document: "Pictorial photography". Kodak says the negatives can be enlarged 25x. For a 24x36 negative, that is about a 24"x36" print. This is about the size one can print with a 24 mpix camera before pixels start showing on fine subjects. If Kodak believes this is about the maximum enlargement possible, I believe them.

      Last be bit least, MTF data from Canon, Nikon or Sony lenses is given at 10 and 30 cycles/mm. MTF data from Zeiss is given at 10, 20 and 40 cycles/mm. No manufacturer of SLR lenses gives data at 80 cycles/mm. If lens manufacturers use these values, I suppose the values are representative.

      Delete
    5. I don't care about the use of tech pan for general photography. That's not the point.

      The point is that tech pan is an actual useful measuring tool for lens resolution, and a DSLR isn't.

      The heroics necessary to actually separate an excellent lens from an OK lens will slam immediately into the limitations of the DSLR.

      Delete
    6. I see.

      This is a sample from a midrange zoom lens, taken at the corner position on a DSLR:

      http://www.slrgear.com/reviews/zsamples/nikon24-120f4G/ynikon24-120f4vfa024f40ul_d3x.jpg

      This is a sample from a very good lens, taken at the same position, also on a DSLR:

      http://www.slrgear.com/reviews/zsamples/zeiss55f14otus/1dsmkiii/VFA/yzeiss55f14vfa055f80ul_1dsmkiii.jpg

      To me, the differences are quite visible, even on a DSLR. Which is the reason why I believe a DSLR is a tool sufficient to determine which of the two lenses is the better one.

      Delete
    7. And you can, after a fashion, slice cheese with a brick.

      Given that the number of variables held constant between those two photos is zero, (different bodies, different focal lengths, zoom versus prime... I guess the test chart is the same) and that the substantive differences are largely chromatic aberration, I think perhaps I am ready to move on.

      Delete
    8. Well, one more minor quibble. That's essentially a field flatness test, not a sharpness test.

      Delete
    9. You must be right:
      -technical pan resolves 4 times what Kodak says
      -all lenses have the same sharpness
      -the poorer lenses resolve twice what the manufacturer says
      -DSLRs are unable to show differences between lenses.

      It is indeed time to move on. Sorry to have bothered you, that won't happen again.

      Delete
    10. Don't let the door hit you on the butt.

      Delete
  2. Finally. Somebody who actually knows the difference between "the math" and arithmetic. Praise the Lord.

    What was the rest of it about ? My eyes slid off the page...

    ReplyDelete