Dot-tune AF adjustment

[dslater]

Has anyone heard of or tried this method of AF adjustment? Does it work? Would you recommend for cameras that don't have the new auto-tune feature?

https://m.youtube.com/watch?v=7zE50jCUPhM

[simsurace]

It is all I ever used in the way of AF adjustment. It has worked fine for me as far as I can tell when lenses need adjustment (and almost all of them have needed it to some degree on my cameras).

There are certainly more precise methods using special equipment. But the precision is maybe not really required when real-world application of AF is fraught with so much imprecision and so many unknowns. The important thing is whether hit rate goes up after adjustment, if it doesn't, just reset the AF fine tune or try something more elaborate.

[longzoom]

Has anyone heard of or tried this method of AF adjustment? Does it work? Would you recommend for cameras that don't have the new auto-tune feature?

https://m.youtube.com/watch?v=7zE50jCUPhM

         I would not follow that, for 2 main reasons.  1 - the AF misbalance is not linear. For ex. my 50/1.8 G SE needs +1 adjustment at minimal distance, and +3 at infinity. My 80-400, at different focus/zoom position needs to be  balanced/adjusted from +10 at 80mm at minimal distance, to +13 at 400mm at infinity, at apertures 5.6 - 8.0.  So I left it at +13.  Second reason - all of lenses, ever produced, have some, less or more, focus shift with closing aperture down. My way - chose the aperture most used for every given lens, make FT adjustment at that aperture. Difference to wide open is not that significant, in most cases, but anyway, if you are perfectionist...   Hope it helped.  Good luck!   LZ

[armando_m]

It is all I ever used in the way of AF adjustment. It has worked fine for me as far as I can tell when lenses need adjustment (and almost all of them have needed it to some degree on my cameras).

There are certainly more precise methods using special equipment. But the precision is maybe not really required when real-world application of AF is fraught with so much imprecision and so many unknowns. The important thing is whether hit rate goes up after adjustment, if it doesn't, just reset the AF fine tune or try something more elaborate.

yes, did it with all my lenses worked great

[David H. Hartman]

It looks interesting to me. I'll give it a shot. I'll recommend doing an a/b comparison with the next AF fine tune value on each side of that determined by this method and evaluating it on a computer monitor.

Dave

Here is a link to the charts I made...

Metric Lens Resolution Charts on Dropbox

This is what the chart looks like...



Enjoy!

[Les Olson]

One problem with this method is that it relies on the assumption that live view focusing is accurate.  Manual focusing in live view using a magnified screen is somewhat more reliable than either the camera's AF or manual focusing in the viewfinder (see Dave Etchells' article http://www.slrgear.com/articles/focus/focus.htm) - but it does not give you critical focus every time and you can't tell when it has and when it hasn't.     

A second problem with the method as it is described in the video - and all the other methods I have seen - is statistical.  There are two kinds of errors in AF systems: an offset, such as a difference between the length of the optical path that the AF system has been given to use in its calculations and the actual length, that leads to a systematic error, and random errors.  AF fine-tune only deals with the system's offset - it creates an equal and opposite offset.  Etchells' data shows that for all focus methods the random errors are bigger than the offset - and that is for a static, high-contrast target - so there is a strong case for not bothering about possible systematic error. 

However, if we are bothering, we have to separate the random and systematic errors and that is not easy.  Because random error can counteract an offset as well as aggravate it, you can't estimate the offset accurately unless you have enough samples to be sure that the random errors in one direction, that counteract the offset, balance the random errors in the other direction, that worsen it.  A back-of-the-envelope calculation using Etchell's data and my comparison of my D7000's AF fine-tune units to centimetres of camera movement suggests that to get an estimate of the offset accurate to +/- 2 fine-tune units you need a lot of trials at every AF fine-tune setting - 30 or more.   

[dslater]

One problem with this method is that it relies on the assumption that live view focusing is accurate.  Manual focusing in live view using a magnified screen is somewhat more reliable than either the camera's AF or manual focusing in the viewfinder (see Dave Etchells' article http://www.slrgear.com/articles/focus/focus.htm) - but it does not give you critical focus every time and you can't tell when it has and when it hasn't.     

A second problem with the method as it is described in the video - and all the other methods I have seen - is statistical.  There are two kinds of errors in AF systems: an offset, such as a difference between the length of the optical path that the AF system has been given to use in its calculations and the actual length, that leads to a systematic error, and random errors.  AF fine-tune only deals with the system's offset - it creates an equal and opposite offset.  Etchells' data shows that for all focus methods the random errors are bigger than the offset - and that is for a static, high-contrast target - so there is a strong case for not bothering about possible systematic error. 

However, if we are bothering, we have to separate the random and systematic errors and that is not easy.  Because random error can counteract an offset as well as aggravate it, you can't estimate the offset accurately unless you have enough samples to be sure that the random errors in one direction, that counteract the offset, balance the random errors in the other direction, that worsen it.  A back-of-the-envelope calculation using Etchell's data and my comparison of my D7000's AF fine-tune units to centimetres of camera movement suggests that to get an estimate of the offset accurate to +/- 2 fine-tune units you need a lot of trials at every AF fine-tune setting - 30 or more.   

Yes, but doesn't the statistical issue apply to other methods of calibration as well?

[Airy]

Indeed, I've been using a special target for AF fine tuning (forgot the brand) with an angled ruler, using liveview etc. but apart for the systematic offsets (relative to distance or aperture-related focus shift) there were random errors, quite significant e.g. with the 50/1.4G, less so with the 85/1.4 or 105VR. So I guess that a quicker procedure should not be rejected a priori.

[Ilkka Nissilä]

Random errors in LV AF are quite large in old cameras (such as the D3X, in my experience) but they seem to be very small in the latest professional models. In repeated LV AF and automatic focus fine tune operations I get a standard deviation of about 1-2 fine tune points with most lenses using the D5 body. I think repeating the operation a few times is advisable but that should be sufficient given adequate quality of focus test target, dark homogeneous background, and proper lighting  on the main target.  I think the same should apply for the dot method if you want to use it (one problem is the dot has in many cases a huge range where it is lit, making the information not so precise).

With the D3X, the dot method does not work well at all because the live view AF operation is so imprecise especially in an indoor setup with limited light available. Thankfully, live view and live view AF have improved a lot since. With the D750, the dot is active in a much larger range than with the D810, so with the D810 camera the dot method gives good starting points for real-world testing; with the D750 I have relied more on real world data than the dot method results because often the dot is active also at +20 or -20 so the midpoint of the range can not be estimated in such cases.

A more rigorous procedure involving taking of multiple shots at each fine tune setting, evaluating the contrast in the ruler area and determining the fine tune value that minimizes the autofocus offset can be used with any camera and it works well also with older cameras. However, the procedure is slow and tedious. Thankfully the newer cameras can be fine tuned in easier ways.

[Les Olson]

Random errors in LV AF are quite large in old cameras (such as the D3X, in my experience) but they seem to be very small in the latest professional models. In repeated LV AF and automatic focus fine tune operations I get a standard deviation of about 1-2 fine tune points with most lenses using the D5 body. I think repeating the operation a few times is advisable but that should be sufficient

If the mean correction is 5 fine-tune units and the standard deviation is 2 and you do 4 repetitions the 95% confidence interval for the mean is 3, 7.  Which sounds OK, but it means that if the correct value is 5, 1 photographer in 20 who does it, or 1 time in 20 you do it, the value you get will be less than 3 or more than 7.  As the man said, "You've got to ask yourself one question. Do I feel lucky?". 

You have to do 50 repetitions before there is less than 1 chance in 20 the correct value is nearer 4 or 6 than 5.

People who only want to do a few repetitions should just be testing a few values - say, 0, 5, 10 and 15. 

[dslater]

Random errors in LV AF are quite large in old cameras (such as the D3X, in my experience) but they seem to be very small in the latest professional models. In repeated LV AF and automatic focus fine tune operations I get a standard deviation of about 1-2 fine tune points with most lenses using the D5 body. I think repeating the operation a few times is advisable but that should be sufficient given adequate quality of focus test target, dark homogeneous background, and proper lighting  on the main target.  I think the same should apply for the dot method if you want to use it (one problem is the dot has in many cases a huge range where it is lit, making the information not so precise).

That is really surprising. I can see why there would be errors in the normal AF system since it uses a different optical path from the actual shot which includes a moving object - namely the mirror. However, when using live-view focusing, you're using the exact same optical path as for the shot, and none of the components are moving.
  So what's the source of variation when using live view focusing? Is the lower resolution of the older camera's live view screen making it difficult to judge when the image is truly in focus? Would using a focusing loupe instead of/in addition zooming in on the live view screen improve this?

[Ilkka Nissilä]

Between 1 and 2, not 2.

If I do four test shots and  the STD of an individual measurement is 1.5, then the standard error of the mean is 0.75. This accuracy is good enough for me especially since the test conditions are not representative of all conditions where the lens and camera combination will be used, and the distance, colour of the light, etc. may affect the optimal value slightly and there is no way I'm going to be able to correct for it all.  So far the values from my auto fine tune protocol on the D5 have worked well and I haven't had to tweak the result based on real-world use on this camera. Even at long distances the autofocus of the D5 is accurate, which is a first in any Nikon that I've used.

With the 24-70 and 70-200/2.8 I get some focal length dependency (+-5 points around the mean) so I use the middle value between the tele and wide ends as the value that I keep in the camera. This still works well enough for the most part but may not be ideal on higher resolution bodies.

[dslater]

If the mean correction is 5 fine-tune units and the standard deviation is 2 and you do 4 repetitions the 95% confidence interval for the mean is 3, 7.  Which sounds OK, but it means that if the correct value is 5, 1 photographer in 20 who does it, or 1 time in 20 you do it, the value you get will be less than 3 or more than 7.  As the man said, "You've got to ask yourself one question. Do I feel lucky?". 

You have to do 50 repetitions before there is less than 1 chance in 20 the correct value is nearer 4 or 6 than 5.

People who only want to do a few repetitions should just be testing a few values - say, 0, 5, 10 and 15.

Well, given this, I'd say that using the method of photographing a scale at a 45 degree angle would be faster than the dot-tune method. You could take 50 shots in a row - defocusing and refocusing in between each shot, then examine the resulting images and determine the correction value for each image, then apply some statistics to determine the mean value as well as the standard deviation of your sample.
  Also, given there are random variations anyway, it seems to me the 95% confidence level is really more that you need. I would think using a sample size of 30 would be enough to get you a statistically significant result.

[Ilkka Nissilä]

That is really surprising. I can see why there would be errors in the normal AF system since it uses a different optical path from the actual shot which includes a moving object - namely the mirror. However, when using live-view focusing, you're using the exact same optical path as for the shot, and none of the components are moving.
  So what's the source of variation when using live view focusing? Is the lower resolution of the older camera's live view screen making it difficult to judge when the image is truly in focus? Would using a focusing loupe instead of/in addition zooming in on the live view screen improve this?

One reason is that the live view image in the D3X is noisy (likely it also doesn't have the full sensor data) and so the LV CDAF autofocus process is subject to this noisy and likely incomplete data. Additionally, most Nikon lenses (SWM) are designed to focus approximately in the correct neighbourhood, going very quickly from distance A to B, rather than going slowly but precisely exactly to the specified distance B. The SWM focusing process is fast but there is some randomness in the end result. The D3X LV AF is a very early implementation by Nikon and may also not use the best CDAF algorithms, and the result is what it is. It is not reproducible enough for use as a reference for autofocus fine tuning, in my experience. Newer cameras such as the D810 are much better in this respect and can tolerate also some low light conditions.

Even Nikon recommends that the user zoom in the image and verify that live view AF focused on the correct position before activating the auto AF Fine tune feature. I think in addition to this recommendation they should state that multiple trials are highly recommended. The difference might not be huge but it's much more trustworthy in my opinion, if one repeats the procedure a few times.

I think it is to be expected that some variation in the optimal fine tune value due to differing conditions may be of the order of 3 points in the fine tuning scale (based on my experiences with various Multi-CAM 3500 cameras; I don't have enough experience with Multi-CAM 20k to say how much the optimal correction depends on conditions using the new system). With some lenses even one point error can be significant though; e.g. 200/2.  When photographing an approaching subject,  I will typically adjust the fine tune setting by -1 or -2 relative to my normal value used for static subjects as I find that it needs a little help to keep track of the subject when it is approaching (I find that this improves my keeper rate).  When I'm done shooting, I return the normal value of fine tune that is optimal for static subjects. But I don't find this procedure to be really needed for f/2.8 and slower lenses; I   use it with a few fast primes.

[dslater]

I'm at work now, so can't play the video again, but I though that when he used live view, he was focusing manually and judging the magnified image on the screen instead of using live-view AF.