Monday, October 19, 2015

Ugly Reality, a Simple Test

Blrrrg. Sorry, there seems to be no way to conveniently zoom in on these pictures, so you're stuck looking at 77 pixel high JPEGs. I thought about going and scaling them all up 4x or something, but I'm exhausted with it all, so please just squint.

This happens to illustrate the uselessness of the 100% crop. You actually need to go to 400% or more to see what the hell is actually going on. All these little clippings look pretty similar, but they're not. If you're super interested in this crap a) God help you and b) download the pictures and look at them much bigger.

I used to be a bit of a lens testing hero, ages and ages back. So, here's a couple tests I ran today, just to give you some flavor. Nikon 300 f/4.5 AIS lens, wide open, happened to be what I had mounted. It's a decent lens, but not brilliant. You can pick them up today for less than $200 (AI) or $300 (AIS) from KEH. Grabbed a book with a nice fine bar chart over the ISBN, tossed it against a wall on the far side of the yard. Careful focusing, and here's a 100% crop for you to look at:

The picture above is updated. Zoomed in to 400%, the red mark indicating, roughly, 2 line-pairs, is 35 pixels long, which yields 17 pixels per line pair, or 4.25 pixels (21.25 microns) per line pair on the sensor, let's call it 22 microns per line pair, for 45 lp/mm. So there.

I have a Nikon D3100, 16 megapixel crop sensor. Calculate a little and you're gonna get a dot pitch of about 5 micrometers. This picture (and all of them in this post) is 77 pixels high, 100% crop. So we're looking at 77x5 = 385 micrometers vertically on the sensor here. Check my work here, but I'm getting 18 maybe 19 line pairs resolved in the vertical direction here, which works out to around 50 lp/mm of resolution, which honestly looks pretty darn usable here. We're starting to flirt with sensor limitations here, but things look pretty good.

Same deal, but the extreme corner of the sensor, about 13mm off the center (crop sensor, sorry). You can see the lens is degrading, but guess what, we still have usable resolution at about 50 lp/mm.

This sensor has the same dot pitch as the D810, and I can challenge it with a $300 lens, wide open.

Let's look at a lens that's actually good. This is likely the best lens I own, run you something like $400 or so at KEH, the 60mm Micro-Nikkor, with a very well deserved reputation. Wide open at f/2.8, 13mm off center. Same 385 micron sensor patch, but a slightly more open frame, and we can really see the wheels starting to fall off. I am seeing high contrast source data beginning to drop into a hole by around 65 lp/mm or so, and it feels to me like the lens losing it slightly before the sensor.

I don't happen to have an Otus lying around, and I dare say it would look a little better. Not, though, a whole lot better, because there simply isn't a lot of better available at the sensor. It doesn't matter how good of a lens you have, the physical limits of the system are staring you right in the face here. Have a look:

Stopping down to f/5.6 and pulling back a little to push the limits, we get this, where the second densest set of lines hitting the sensor around 70 lp/mm are almost gone, and the densest, hitting around 90 lp/mm, are just slightly modulated mush. These are hard sensor limits here. The lens is pretty much out of the picture. You could spend a million dollars on a lens and get no more resolution out of it.

This one is also expanded 400%. I also overlaid a strip of a much sharper picture of the same bar chart, for reference, so you too can see what the bars actually look like. The blue circle indicates a couple lines spaced about 13 pixels apart (3.25 pixels, 16.25 microns at the sensor) and the red lines indicate lines spaced about 9 pixels apart (2.25 pixels, 11.25 microns) at the sensor.

In summary, the D810 is useless as a test horse for lens resolution measurements, except for cheap lenses, wide open, and measured in the corners.


  1. Arithmetics again and then I am gone, because this is indeed enough about the subject.

    The D3100 has a 23.1 x 15.4 mm sensor and a resolution of 4608 x 3072. 4608 divided by 23.1 gives about 200 pixels per mm, which translates to 100 cycles per mm (2 pixels per cycle, extinction resolution).

    Looking at your pictures, the finer black lines are about 3 pixels wide, so you are testing 100/3, or about 30 cycles/mm.

    1. You consistently seen to think I am an idiot, which I find tedious. You begin by repeating in different words the statement about 5 micron dot pitch which I made, which suggests to me you haven't bothered to read what I wrote. This isn't the first time you've said such things.

      Then you have the temerity to accuse me of muddling up cycles and dots and a moment later talk about line widths versus line spacings.

      Do you mean that in one of the four pictures at three different magnifications, you saw some lines spaced at three pixel intervals? Or is the line width 3 pixels, implying a line spacing of 6 pixels? Or what?

      Line widths are meaningless in this context. You have to measure peak to peak, ideally across a group of lines.

      I'm certainly not going to write a lecture on his to read a test chart here, but if you don't know how, I can going that you ought to go find out.

  2. I see a line width of 3 pixels, implying a cycle width of 6 pixels peak to peak on the finer structures, yes.