[Lev]

Well guys, testing this lens these days.. pushing it to its limits and shooting at wide apertures. Yes you must be very careful shooting at 1.4 but I have one question regarding autofocus accuracy.

Yesterday I've done some tests with the focus chart and shot at 1.4 from about 60cm distance from target. 2 or 3 shot out of 10 were slightly off. I remember when I first bought 50-135 had veeery bad results, had terrible front/back focusing. Well, this lens shows very little focus shift, it's not front or back focusing issue, it just sometimes little off towards front or back. I have checked my DFA 100mm macro and it also has a little deviation but it's two stops slower and 0 point stays sharp during shifts. Now I'm very interested, is it cause of phase detect autofocus system, body, or lenses? In live view star lens focuses with 100% accuracy.

[stevebrot]

Yes you must be very careful shooting at 1.4 but I have one question regarding autofocus accuracy.

The answer is complicated, but is not related to accuracy. What you are seeing is the precision or lack of such due of the characteristics of the PDAF system. As you may be aware, the K-1 has two classes of AF sites, those sensitive to f/5.6 and those sensitive to f/2.8. By "sensitive", I mean the ability to detect being out-of-focus. If PDAF is sensitive to f/5.6, it means that the ability to detect OOF is the same looking through a f/1.4 aperture as through an aperture of f/5.6. The result is lower precision than with the same lens using one of the f/2.8 AF sites, but still not as good as might be possible if the K-1 offered sites sensitive to f/1.4*. Strangely, the narrower depth of field as well as more light to the sensor does not result in better focus precision (fewer images that are almost in focus). You will get a similar number of shots in perfect focus as might be available with an f/2.8 maximum aperture lens with the same subject. Of course, this behavior would be most obvious over the course of many exposures.

The reasons are related to how PDAF works and the optical limitations of split-image rangefinders. The full explanation requires perhaps a thousand words and many diagrams**. As you noted, CDAF in live view tends to have better observable precision, partially because its sensitivity is similar to that of the person viewing. When it falters, we simply don't notice. While the PDAF site sensitivity is a matter of physics and the sensor's design, there are other factors that make perfect precision difficult. Those include lens contrast, nature of the subject, and mechanical imprecision in the focus mechanism.

I might add that similar issues are traditionally present even when using manual focus. The stock focus screen on most dSLRs has a sensitivity of about f/4.0. The split image on aftermarket screens is sensitive to about f/1.4 at best and the highly regarded Canon S-type screen is sensitive to f/1.8 but is problematic at apertures narrower than f/2.8.


Steve

* Not offered on any commercial product that I am aware of.

** For a detailed explanation see Doug Kerr's excellent article on how PDAF works and its relationship to split-image rangefinders: Link to PDF

[pschlute]

2 or 3 shot out of 10 were slightly off

That sounds to me about the correct hit rate for PDAF systems.

[Lev]

Steve, so on K-1 precision should be better right? And also I assume that your answers are based on fact that my tests were done in controlled environment with testing chart on tripod...

[FozzFoster]

For a detailed explanation see Doug Kerr's excellent article on how PDAF works and its relationship to split-image rangefinders: Link to PDF

This is a great read thanks for sharing!

[zapp]

The correction value should in theory always be around zero for all lenses. My lenses have perfromed the same on each of my cameras. Sometimes one camera was off by a couple steps from zero and so were all the lenses. The differences between lenses in relative terms were exactly the same for all cameras. Trouble starts only in cases where the camera offset is too large to compensate for additional lens offset. Allow +/-3 to 5 values for the camera and the same margin for the lenses and you can hit the extreme 10 easily. In such cases the camera can be re-zeroed.
Focus shift can depend on focus distance and for zooms also on focal length. The 1.4/50 was not that sensitive with focus distance if I remember correctly, but I corrected my wide open calibration for portrait range. My 70-200 has different optimum values for 70/135/200 zoom settings, to I had to decide for a single setting that works best overall.
Always remember it is not only the lens that has an offset, also the camera focus system can be slightly off. I always use lens calibration and check from time to time. When my K1ii came back from intesnive camera care, it was perfectly calibrated for the 1.4/50 that went with it in intensive care. Knowing the new calibration setting for one of my lenses, it was easy to adjust all the others from previous tests (double checked).

[UncleVanya]

Steve, so on K-1 precision should be better right? And also I assume that your answers are based on fact that my tests were done in controlled environment with testing chart on tripod...

Not that I'm aware of. F2.8 sensitive af points exist in many bodies including the k-3 - the k-1 doesn't have anything different. Better brains for tricky focus situations perhaps but not really different precision.

[Lev]

Focus shift can depend on focus distance and for zooms also on focal length

In my tests, I haven't changed focus distance, I was just refocusing on every shot. Thing is that the lens doesn't have constant front or back focus problem, it just sometimes slightly misfocuses. I have just adjusted microfocus value to -2 because I'd like to have areas in focus just little bit behind subject rather that in front. Mostly it nails focus just perfectly, absolutely sharp at zero. I just started worrying when it misfocused in controlled situation on tripod, at the same target. I think I'm asking too much from PDAF Steve? Or it should be perfect on tripod?

[stevebrot]

Steve, so on K-1 precision should be better right? And also I assume that your answers are based on fact that my tests were done in controlled environment with testing chart on tripod...

I can't really say. I guess it depends on one's personal expectations, though I might hint that even the best from Canon or Nikon will perform similarly if tested with the same lens (e.g. the Tokina Opera 50/1.4). I personally use manual focus in magnified live view when critical focus is the goal.

I think I'm asking too much from PDAF Steve?

Yes


Steve

[stevebrot]

The correction value should in theory always be around zero for all lenses.

Yes, if the lenses achieve theoretical perfection and the subject has infinite contrast. The need for AF fine tuning is due to asymmetry or ambiguity of the image provided to the sensor site. In regards to the original post, the issue is probably not focus shift or problems with fine tuning; rather, it looks to be the expected scatter due to probability.


Steve

[Paul the Sunman]

In regards to the original post, the issue is probably not focus shift or problems with fine tuning; rather, it looks to be the expected scatter due to probability.

So, if as you are saying, the errors are stochastic rather than deterministic, does that mean that different realisations will return a range of outcomes within that band? In particular, if I repeatedly focus on a subject by moving my finger on and off the shutter release button, will the focus just randomly shift by an amount of the order of the standard deviation, or will it more-or-less monotonically improve?

[Lev]

it looks to be the expected scatter due to probability.

So I wonder if this probability is legit even when the camera and the subject are steady and AF point aims at same place on target surface shot by shot....

---------- Post added 12-18-18 at 01:38 PM ----------

So, if as you are saying, the errors are stochastic rather than deterministic, does that mean that different realisations will return a range of outcomes within that band? In particular, if I repeatedly focus on a subject by moving my finger on and off the shutter release button, will the focus just randomly shift by an amount of the order of the standard deviation, or will it more-or-less monotonically improve?

good question. As I know, focus doesn't improve by repeatedly pressing focus button, it just doing it again imho.

[stevebrot]

In particular, if I repeatedly focus on a subject by moving my finger on and off the shutter release button, will the focus just randomly shift by an amount of the order of the standard deviation, or will it more-or-less monotonically improve?

That is a very sticky question that I don't have the time to dissect completely. I guess one would have to do some empirical testing to see. At the very least, I would expect that any incremental improvement would be equally successful for an f/2.8 lens as for an f/1.4 lens.


Steve

[Paul the Sunman]

That is a very sticky question that I don't have the time to dissect completely. I guess one would have to do some empirical testing to see. At the very least, I would expect that any incremental improvement would be equally successful for an f/2.8 lens as for an f/1.4 lens.
Steve

My experience is that refocussing once often improves focus accuracy. But after that I don't know. It may well descend into a tight random well.

[stevebrot]

So I wonder if this probability is legit even when the camera and the subject are steady and AF point aims at same place on target surface shot by shot....

What I suggested is the known element of imprecision as far as detection is concerned. Do the comparison again with the D FA 50 set to f/2.8 and the D FA 100 wide open to allow similar DOF to affect the final results. Remember to adjust the distance for the D FA 100 to provide equivalent magnification. The D FA 100 will be at a slight disadvantage, but I would expect AF precision to be similar.


Steve