I’ve been reporting on the results of testing sharpness, as measured by MTF50 results from slanted-edge testing with the Otus 55/1.4 on the Nikon D810 and the Sony alpha 7R (a7R). For the last few days, I’ve posted nothing. That hasn’t been because I’ve stopped working. It’s been because I haven’t been able to produce reliable, repeatable, accurate results.
First off, there’s the inherent statistics of the camera and measurements. Take 16 pictures as I do, under the same conditions, and you’ll get 16 different results. 16 pictures is enough to get fairly accurate mean results, but you can’t cut down the number of pictures per data point without introducing the possibilities of chance materially skewing the results. The spreads were pretty tight with the strobe-illuminated images, but they are looser with continuous illumination.
The second thing is that focusing needs to be incredibly accurate to get meaningful comparisons at the Otus’s sharpest f-stop, which is about f/4. In fact, I have not yet devised a way to get repeatable results with the D810, using live view at f/1.4 and maximum magnification to focus on the Siemens star on the target.
It’s not a problem with the a7R, not so much because of the better live view resolution, but due to focus peaking, which is unavailable on the D810. Peaking works great on the Siemens star, and provides a microscope into whether the image is at its sharpest.
I’m not sure what to do about the focusing issue. One possibility is to look for other focusing targets, and print out an Imatest target with the Siemens star replaced by one that works better on the D810, In the past, I’ve had better luck focusing on the zone plate at the center of the ISO 12233 target than with the Siemens star.
Whether I focus on the star or a zone plate, in the absence of focus peaking, my strategy is to concentrate on an appropriate aliasing artifact, and find the focus point that maximizes the artifact. Kind of a perverse way to focus, but it works well.
Another possibility is to get a big display, connect it to the D810 with an HDMI cable, and use that for focusing.
Yet another potential solution, which I categorically reject, is to refocus each exposure. There are usually more than 200 photographs in a series, and it’s maddeningly boring already — though less so since I realized I can use the D810’s built-in intervalometer to do each set of 16 exposures.
Trying to trace what’s behind the non-repeatibility of some of he measurements I’ve been making (and won’t show you; they confuse even me, and I made them) has caused me to examine many of the plots that Imatest does for each analysis. That has been instructive, if time-consuming. I have found evidence of camera motion on both the D810 and a7R images, even in situations where you’d expect such motion to be minimal. Let me show you a few.
This image was made with the D810 in landscape orientation, looking at a (highlighted) vertical edge.
Now let’s look at a horizontal edge of the same exposure, which will suffer more from the second curtain of the shutter (the first curtain is not being used) since the shutter moves up and down on the D810:
See that little notch in the MTF curve, just above the Nyquist frequency? That looks for all the world like the effect of a mild anti-aliasing (AA) filter. But neither camera has such a filter, and this effect doesn’t occur in all the images from either camera.
Even some of the pictures that were made with the strobe show the effects of camera vibration, showing that the short-duration Einstein isn’t short enough to allow us to stop worrying about camera motion. Here’s the vertical edge from one such image, this time with the a7R:
And here’s the horizontal edge from the very same exposure:
If I didn’t know better, and hadn’t seen pictures from the a7R that don’t exhibit this effect, I’d swear it had a mild anisotropic AA filter.
It’s all very frustrating. Yet, I persist, since being able to make consistent, repeatable images of test charts is instructive as to ways to make consistent, repeatable images of the real world.