Auto-focus sensors: high precision vs. regular testing

But wouldn't any focus shift impact the absolute focus point, and not the focus accuracy? The shift is always to the same place and in the same direction as a function of aperature, is it not? It is not a random statistically signficant shift like what he is trying to measure here. I think he is trying to get at 'accuracy' and 'repeatability', not the absolute focus point.

Assume it does shift, but if the shift happens the same way each and every time as a function of aperature, then it doesn't affect the repeatability or the statistics. That's what I was trying to get at.
I have no evidence, only one of many theories, but perhaps people are looking at these things more closely, and things that haven't been noticed, or were imperceptible w/lower resolutions, are now being discovered.

Certainly the 50L has a lot of focus shift. So do other fast lenses, like the Noctilux which is known for its quite severe focus shift. Some enlarging lenses ( EL Nikkor 50mm f2.8, for one)do have a focus shift when changing aperture.

It wouldn't be a shock if slower lens have some slight focus shift. We are talking about a very sensitive test here.
--
kind regards
Dale
 
Assume it does shift, but if the shift happens the same way each and every time as a function of aperature, then it doesn't affect the repeatability or the statistics. That's what I was trying to get at.
And you are right. I do not see focus shift as being relevent here for the reasons you have already stated. Additionally, I tested the 70-200 f2.8 II wide open (only changing the selected AF point) and the focus variation was still there. In fact, at wide apertures is when the variation will be most apparent in the final image.

I am not surprised there is focus variation, the interesting part is that the Center (high precision) AF point only had less focus variation under high light levels. At lower light levels it had more focus variation.

The reason this is significant is so many people had assumed that f2.8 and better lenses would have better AF in low light. That is clearly not the case, even for the newest APS-C body and lens.
I have no evidence, only one of many theories, but perhaps people are looking at these things more closely, and things that haven't been noticed, or were imperceptible w/lower resolutions, are now being discovered.

Certainly the 50L has a lot of focus shift. So do other fast lenses, like the Noctilux which is known for its quite severe focus shift. Some enlarging lenses ( EL Nikkor 50mm f2.8, for one)do have a focus shift when changing aperture.

It wouldn't be a shock if slower lens have some slight focus shift. We are talking about a very sensitive test here.
--
kind regards
Dale
--
Mike Mullen
 
Yes... I understood all of this from reading your data and other posts, mmullen. I think this is groundbreaking work. I have no explanation for what you are seeing, but it rings true with my own observations.

I have noticed that I didn't seem to be getting greater precision with 2.8 and faster lenses -- but never knew why or made an attempt to pin it down as you are doing.

As I said, groundbreaking work. You've taken the first steps towards finding understanding. Bravo!

The only suggestion here I might offer is to try with targets of larger detail and greater contrast than towel patterns. I suspect that the precision sensors are also courser in that they don't do as well with textures as they do with large, welll defined, high contrast edges... Just a gut feeling...
Assume it does shift, but if the shift happens the same way each and every time as a function of aperature, then it doesn't affect the repeatability or the statistics. That's what I was trying to get at.
And you are right. I do not see focus shift as being relevent here for the reasons you have already stated. Additionally, I tested the 70-200 f2.8 II wide open (only changing the selected AF point) and the focus variation was still there. In fact, at wide apertures is when the variation will be most apparent in the final image.

I am not surprised there is focus variation, the interesting part is that the Center (high precision) AF point only had less focus variation under high light levels. At lower light levels it had more focus variation.

The reason this is significant is so many people had assumed that f2.8 and better lenses would have better AF in low light. That is clearly not the case, even for the newest APS-C body and lens.
I have no evidence, only one of many theories, but perhaps people are looking at these things more closely, and things that haven't been noticed, or were imperceptible w/lower resolutions, are now being discovered.

Certainly the 50L has a lot of focus shift. So do other fast lenses, like the Noctilux which is known for its quite severe focus shift. Some enlarging lenses ( EL Nikkor 50mm f2.8, for one)do have a focus shift when changing aperture.

It wouldn't be a shock if slower lens have some slight focus shift. We are talking about a very sensitive test here.
--
kind regards
Dale
--
Mike Mullen
--
kind regards
Dale
 
My understanding is that a lens of f2.8 or faster always starts with a regular precision AF sensor and then when focus is almost achieved the high precision AF sensor is given a chance to take a reading and fine-tune the focus position.
Mike, do you remember where you might have come across that information? I've looked quite a bit on the internet for some sort of "official" technical explanation for that very thing, and come up dry.
I'm sorry, I don't. I do recall it was a statement directly from a Canon rep.
Mike, I've found an online reference to that information -- in the 5D White Paper, p. 9. There's probably lots of other sources as well, but now that I've found this one, I doubt I'll keep looking for others. :)
But the results of your own tests seem to contradict that notion. In other words, if the center AF point really did work that way (focus first using f/5.6 sensor, then pass off to f/2.8 sensor for fine-tuning), the result would never be worse than using just an f/5.6 sensor alone -- possibly better, but never worse.
No. If the f2.8 sensor didn't have enough light to work reliably, it could cause the focus to be "touched up" in the wrong direction.
Yes, it appears to be exactly as you've said ( if the 7D's logic is the same as the 5D's, that is). Here's the pertinent quote from the 5D White Paper noted above:
"The center AF point has a special hybrid design. With f/2.8 or faster lenses, focusing is a two-step process. First, the f/5.6-sensitive cross-type sensor components are used to focus. When focus is almost achieved, a switch is made to the f/2.8-sensitive vertical line-sensitive sensors for high-precision focusing."
My mistake was in assuming that if it worked that way, the camera would at least try do the best job it possibly could with the f/5.6 sensor in the center AF point, before passing off to the f/2.8-enabled sensor for focus refinement. But that "almost" in the explanation of its actual operation puts that thought to bed. And even if the f/5.6 sensor did manage to nail focus perfectly before passing off to the f/2.8 sensor, then -- as you said -- the f/2.8 sensor could still make an erroneous adjustment to that setting if it didn't have enough light to work reliably.
--
Greg
 
This thread makes me sad :( I too have noticed I'm getting consistently unreliable autofocus with fast lenses (keeper rate 20% or less), but when I switch to my F4.5 lenses the autofocus is spot on and my keeper rate shoots up to about 80%.

Hope this is something that can be fixed in firmware
 
Please check your AI Servo AF Tracking Method (C.Fn III-3). 7D Manual, page 210.

Do you use:

0: Main Focus Point Priority (or)
1: Continuous AF Track Priority

When focus tracking a subject in AI Servo and a closer subject appears...

Option 0: The active AF point will switch to the main focus point and start focusing the closer subject. So if you use single point expanded and the adjacent point had focus, the main point would take over and start focusing. This is useful only if you want to track the closer subject.

If you shoot with a 2.8 lens the focus will switch faster versus a 4.0 lens.

However if you want to continue tracking your main subject, you should use Option 1.

Let me know what you think.
 
This is interesting. I've found a similar result

It's as if it puts priority on firing the shutter instead of confirming focus completely. This happens 5% of the time for me.
Do you shoot AI Servo..

Check your customer function C.Fn III-2. Do you have it set to #2? This option provides shutter release priority over focusing the subject.
 
Please check your AI Servo AF Tracking Method (C.Fn III-3). 7D Manual, page 210.

Do you use:

0: Main Focus Point Priority (or)
1: Continuous AF Track Priority
This function only comes into play if you are using multiple AF points in Servo AF. I know your comments were not directed at me but for clarity I'll state all of my tests were done in One Shot AF mode using a single AF point.
When focus tracking a subject in AI Servo and a closer subject appears...

Option 0: The active AF point will switch to the main focus point and start focusing the closer subject. So if you use single point expanded and the adjacent point had focus, the main point would take over and start focusing. This is useful only if you want to track the closer subject.

If you shoot with a 2.8 lens the focus will switch faster versus a 4.0 lens.
What is the basis of that claim?

--
Mike Mullen
 
Manually selecting a LP focus point and then a HP one might work or possibly setting up the focus 'target' so that there is only contrast under the sensor under test (with a 2.8 or faster lens attached will it use the HP sensor only and give up if the HP one does not work or will it revert to a LP sensor for a try?)
My second series of tests used the first method (comparing the focus consistency of a manually selected HP AF point with the accuracy of a manually selected regular AF point).

--
Mike Mullen
 
I am too having low light inconsistencies with my 17-55 f2.8 IS and 70-200 f2.8 IS with my 7D but not with my 50D so this thread could explain why. I tried another 7D with my lenses and it is also inconsistent.
 
Please check your AI Servo AF Tracking Method (C.Fn III-3). 7D Manual, page 210.

Do you use:

0: Main Focus Point Priority (or)
1: Continuous AF Track Priority
This function only comes into play if you are using multiple AF points in Servo AF. I know your comments were not directed at me but for clarity I'll state all of my tests were done in One Shot AF mode using a single AF point.
Correct, my comments were not directed at you. I was making a suggestion based on shooting "real" subject and not test subjects like bath towels. In "real" world shooting this option could cause focusing issues depending on what you want to track.
When focus tracking a subject in AI Servo and a closer subject appears...

Option 0: The active AF point will switch to the main focus point and start focusing the closer subject. So if you use single point expanded and the adjacent point had focus, the main point would take over and start focusing. This is useful only if you want to track the closer subject.

If you shoot with a 2.8 lens the focus will switch faster versus a 4.0 lens.
What is the basis of that claim?
Sorry, I did not intend this to be a "claim" but an assumption that a 2.8 lens may switch to the adjacent point faster than a 4.0 lens. I have no proof that that is correct and was looking for feedback.
--
Mike Mullen
Very interesting thread and I appreciate all the work and thoughts. I'm trying to understand if there is any "real world"data to support what appears to be a growing "assumption" that 2.8 lenses are not as "consistant" as slower glass. I'd like to see some data to better understand.
 
Sorry, I did not intend this to be a "claim" but an assumption that a 2.8 lens may switch to the adjacent point faster than a 4.0 lens. I have no proof that that is correct and was looking for feedback.
I've never seen anything to indicate that's true and your statement that 2.8 would switch faster just came out of the blue as far as I could tell.
I'm trying to understand if there is any "real world"data to support what appears to be a growing "assumption" that 2.8 lenses are not as "consistant" as slower glass. I'd like to see some data to better understand.
That wasn't my conclusion. My testing showed that the center point high precision AF point had more consistent AF under good light but less consistent AF under low light conditions. I only thoroughly tested this on a 7D body.

I do not understand your fixation with "real world". In my opinion the testing I did occured in the "real world". I used conditions that isolated the One Shot AF locking onto a planar surface to insure the results were not due to the AF locking onto different subject distances. I see no reason why this would not be a valid way to measure the accuracy of the AF sensors.

I did not test the focus accuracy in Servo AF mode but I have every reason to believe that the high precision AF sensors are NOT used in Servo AF mode (even if the lens is f 2.8 or faster).

--
Mike Mullen
 
I did not test the focus accuracy in Servo AF mode but I have every reason to believe that the high precision AF sensors are NOT used in Servo AF mode (even if the lens is f 2.8 or faster).
That's a very interesting statement. If you've mentioned that before, I must've missed it. Do you believe that's true just for the 7D specifically? Because I'm pretty sure that's not the case, in general, for all models ...

For example, in the 1-Series DSLRs, the cross-type AF points wouldn't even be cross-type AF points, if not for the horizontal (vertical line-sensitive, that is) f/2.8 sensor that's contained in each of those points. Canon notes specifically in their literature that these AF points become simple line-type sensors with lenses of f/5.6-or-slower, but don't mention anything about them losing their cross-type (or f/2.8) capability when the camera is set to AI Servo AF mode.

Furthermore, two of the original 5D's six non-selectable AF assist points incorporate f/2.8 sensors, and those assist points aren't even enabled unless the camera is set to AI Servo AF mode.

Since many of the common uses for Canon's fast telephoto lenses would be in applications to which AI Servo AF is best suited, it seems odd that Canon would deliberately disable the AF sensor(s) that allow for additional AF precision when those lenses are used with the camera in that mode. What is it about the camera's operation that has led you to believe that the f/2.8 sensors are disabled in AI Servo AF mode?
--
Greg
 
Do you believe that's true just for the 7D specifically? Because I'm pretty sure that's not the case, in general, for all models ...
Yes, the 7D and also the 40D. I would not want to comment on models I don't have first hand experience with.

--
Mike Mullen
 
Do you believe that's true just for the 7D specifically? Because I'm pretty sure that's not the case, in general, for all models ...
Yes, the 7D and also the 40D. I would not want to comment on models I don't have first hand experience with.
Understood, thanks. And I'm not saying you're wrong, in regards to the two models you have. But, I mean, how do you know the f/2.8 sensor isn't working with AI Servo on those cameras? What sort of indication are you getting?
--
Greg
 
That wasn't my conclusion. My testing showed that the center point high precision AF point had more consistent AF under good light but less consistent AF under low light conditions. I only thoroughly tested this on a 7D body.
Interesting. I did some un-scientific testing today taking pics of my black lab puppy indoors. In low light it is a difficult AF test because the dog is all black making it difficult to find contrast for AF. I tried focusing on her eyes using one-shot AF and cycled through the available options: spot, single, single point expansion (using both the center and outer points, and zone. The sharpest pictures were always with the spot.

If it makes any difference I was using a 70-200 2.8 with IS on. I will try it with some other lenses on another day.

Not a very "scientific" test but very practical for the type of real world pictures I take.
I did not test the focus accuracy in Servo AF mode but I have every reason to believe that the high precision AF sensors are NOT used in Servo AF mode (even if the lens is f 2.8 or faster).

--
I have no idea if this is true but would be surprised (and disappointed) if this was the case.

I'm looking forward to more feedback because I have lots to learn.
 
But, I mean, how do you know the f/2.8 sensor isn't working with AI Servo on those cameras? What sort of indication are you getting?
I use Servo AF a lot for black dogs running towards the camera and also for BIF's, motorcycles coming towards the camera and the like.

There are two reasons I do not think the high precision sensors are used during Servo AF. The first reason is just my conjecture but the second reason is based upon real world usage:

1) Servo AF takes multiple distance measurements during tracking and uses these measurements to determine at what rate the subject is approaching the camera and whether the subject is accelerating or slowing. It uses those measurements to guess where the subject will be during the next time the shutter open. It is my conjecture that the errors involved in guessing where the subject will be during the next exposure are large enough that they would dwarf any additional accuracy that may be squeezed from the high precision sensor. Additionally, the high precision sensors are not as reliable in getting a reading as the regular sensors (even on static objects, as my testing shows, let alone moving ones) and Servo AF requires a constant stream of good data to provide reliable tracking for a subject that may be constantly changing the rate of closure. In other words, it would be more important to have reliably good readings of a slightly lower accuracy than to have readings that varied from excellent to poor. Remember, the AF system has no way to tell the difference between a high precision reading outlier or a change in speed or the angle of closure of the subject.

2) I have used my 70-200 f4 IS lens extensively in Servo AF mode as well as my 300mm f2.8 and, recently, the new 70-200 f2.8 IS II. I use them on the same types of subjects in the same conditions using the same body and the same AF settings. A lot of my Servo AF usage is of dogs running toqwards the camera. If f2.8 lenses had an accuracy advantage in Servo AF mode it should show up in my images. But it doesn't. So, if there is any benefit, it is so small as to being meaningless. Actually, my sense is that I get a slightly lower hit rate with my top quality f2.8 lenses than with the 70-200 f4 IS which leads me to think larger aperture lenses may be somewhat disadvantaged slightly in terms of reliable Servo AF tracking (and also in terms of AF lock in One Shot AF mode). But it is certainly no better in my experience.

--
Mike Mullen
 
But, I mean, how do you know the f/2.8 sensor isn't working with AI Servo on those cameras? What sort of indication are you getting?
I use Servo AF a lot for black dogs running towards the camera and also for BIF's, motorcycles coming towards the camera and the like.

There are two reasons I do not think the high precision sensors are used during Servo AF. The first reason is just my conjecture but the second reason is based upon real world usage:

1) Servo AF takes multiple distance measurements during tracking and uses these measurements to determine at what rate the subject is approaching the camera and whether the subject is accelerating or slowing. It uses those measurements to guess where the subject will be during the next time the shutter open. It is my conjecture that the errors involved in guessing where the subject will be during the next exposure are large enough that they would dwarf any additional accuracy that may be squeezed from the high precision sensor. Additionally, the high precision sensors are not as reliable in getting a reading as the regular sensors (even on static objects, as my testing shows, let alone moving ones) and Servo AF requires a constant stream of good data to provide reliable tracking for a subject that may be constantly changing the rate of closure. In other words, it would be more important to have reliably good readings of a slightly lower accuracy than to have readings that varied from excellent to poor. Remember, the AF system has no way to tell the difference between a high precision reading outlier or a change in speed or the angle of closure of the subject.

2) I have used my 70-200 f4 IS lens extensively in Servo AF mode as well as my 300mm f2.8 and, recently, the new 70-200 f2.8 IS II. I use them on the same types of subjects in the same conditions using the same body and the same AF settings. A lot of my Servo AF usage is of dogs running toqwards the camera. If f2.8 lenses had an accuracy advantage in Servo AF mode it should show up in my images. But it doesn't. So, if there is any benefit, it is so small as to being meaningless. Actually, my sense is that I get a slightly lower hit rate with my top quality f2.8 lenses than with the 70-200 f4 IS which leads me to think larger aperture lenses may be somewhat disadvantaged slightly in terms of reliable Servo AF tracking (and also in terms of AF lock in One Shot AF mode). But it is certainly no better in my experience.

--
Mike Mullen
Thanks for the reply, Mike. That all sounds reasonable.
--
Greg
 
Since you can select any AF point you want, and only the center is high precision, why not run the 85MM with any point but the center? If the focus improves, it's the HP point and the subject/lighting combination. If it doesn't improve, it's the basic lens characteristic. No need to guess.
 
Brilliant! This has been bedeviling us. Superb solution. 1000 Thanks.
 

Keyboard shortcuts

Back
Top