Open-loop or Closed-loop Phase Detect Autofocus?

WilbaW · Apr 22, 2009

Erik Magnuson said:
WilbaW said:

I understand the argument, but without knowing a lot more about what
lens calibration actually does,

Click to expand...

Depends on the lens and the issue. For modern lenses, it's
re-programming some tables in the lens EEPROM. For others it may be
physical shims or soldering bridges over jumper points (as someone
found in the the 50mm f/1.8 Mk I service manual).

That's how you do it, not what it does.

The question is, what do those methods of adjustment achieve - what aspects of the lens's performance are being altered and in what way?

Member said:
Member said:

You've lost me here - other to what?

Click to expand...

Other behavior for a closed loop system for the case where the lens
would not move to the correct focus point ...

It would not move to the correct focus point because ...?

Member said:
... (i.e., for closed loop, the end states would either be correct focus
or lens hunting. ...

Or an acceptable focus, e.g. "within the DOF". All the system needs is an idea of an acceptable phase difference. Once that's achieved or bettered, it can stop trying.

Member said:
... FF or BF with focus confirmation would not be an end state.)

Does not compute. I guess it depends on how "the lens would not move to the correct focus point".

Member said:
Member said:

I think the key is in how the body relies on the lens to do what it's
told.

Click to expand...

Only the same old sources.

Do you have references? Thanks.

Andreas Helke · Apr 23, 2009

WilbaW said:
Andreas Helke said:

The autofocus system is a mix of open loop and closed loop system.
The first step is a closed loop operation mode to look for the
aproximate focus. If the lens is already reasonably close to correct
focus this step is skipped.

Click to expand...

So you would say that the experiment in the OP only relates to that
step?

Ir may well be. I could not repeat the experiment because my lenses all focus too fast.

WilbaW said:
WilbaW said:

The final focus measurement is open loop. The camera measures the
distance. Very precise is the equipment is calibrated correctly and
you have a high contrast edge to focus on but often seriously wrong
for low contrast subjects. After that open loop distance measurement
step it uses a closed loop again for the actual focusing.

Click to expand...

I often see this stated - is there a definitive source?

The information comes from Canon technical literature. From ovserving how my camera behaves and misbehaves I believe the description is correct.

WilbaW said:
WilbaW said:

It gives the command to focus to the measured distance than asks the lens
if it got it right and orders a correction if the lens reports a move to
the wrong distance. This just checks if the mechanical lens movement
is right. It does not care if the distance the autofocus measured is
correct.

Click to expand...

The way you describe it, there is no loop (it's just a series of
instructions), and the AF sensor isn't involved, so you can't call
that a system closed-loop (seriously!).

I called the lens adjustment closed loop because it relies on feedback from measurements in the lens itself. But its just done once and not until the system stabilizes somewhere. So its not a normal feedback driven closed looped operation. This operation explains why you can often notice an extra lens movement if you have a micromotor lens.

Andreas Helke · Apr 23, 2009

SNGX1327 said:
WilbaW said:

Paul De Bra said:

An easy way to see what happens is to focus in very low light (with a
non-USM Canon lens). In low light everything happens a bit more
slowly, that's why it helps. The focus is first adjusted "rougly",
then better and better... and when the error is small enough the AF
system gives the lens one last command which should achieve perfect
focus.

Click to expand...

Click to expand...

yes, and no.

in normal light, if the AF sensor can easily detect focus, it only
moves the lens once.

in poor lighting, sometimes the AF sensor cannot easily detect focus.
in this situation, the lens will "hunt" for something to focus on.

for instance, in good lighting with high contrast, if you are trying
to focus near but the lens is set to infinity, the lens can still
easily detect where to focus.

in low light in the same situation, the lens has to adjust it's focus
nearer to the subject. instead of a big blur, the object starts to
take shape and contrast starts to appear. at this point the system
may OR may not be able to obtain focus.

this is what happens when you cover the lens with your hand (or leave
the lens cap on). the lens can't find focus, but it "hunts" the lens
in and out in order to see if it can't get a better look at the scene.

another example of this is that in low lighting, if you are focused
too far away from your subject (example focusing on a near subject,
lens set to focus at infinity) the lens might not find AF. it might
just hunt right past your subject. this is especially evident in
telephoto lenses with a shallow DOF.

however you can assist your lens by manually focusing on your subject
so that the subject appears sharp over your AF point. then when you
activate AF, it will be less likely to hunt, more likely to detect
contrast in the subject and lock AF.

SNGX1327 said:

I fundamentally disagree with this on principle.

I can't imagine any engineer designing a high-precision control
system in which the final step is a movement rather than a
measurement. The system only knows that it has achieved it's goal
after measuring the error and judging it to be acceptable. Just
assuming that an instruction was performed perfectly doesn't make any
sense, especially when you can easily check the actual error.

Click to expand...

in PRINCIPLE, yeah it seems faulty. but that's actually how it works.
IF the system measured as its final step then "backfocus" and
"frontfocus" would not exist. the lens would continually adjust the
lens until the image is sharp.

in BF and FF situations, the lens is still producing a soft image
despite the fact that the AF sensor thinks it's sharp. if the AF
sensor measured again, it would realize this error and correct it.

also think about this --- in many older lenses (non-USM), there is a
higher percentage of error when the focus element travels large
distances than short distances. for instance, let's say that the
focus element of a lens rotates 90 degrees about the barrel from
infinity to its nearest MFD.

if you focus on an object, and the camera detects AF, and the lens
travels 70 degrees from its position to obtain correct focus, there
is a greater amount of error than if the lens only had to travel 20
degrees. i wish i had a link and a source for this, but i don't.

this is probably a combination of the AF sensor going off a
less-contrasty out of focus image, and the AF system being less
accurate when moving the focus element long distances.

but we are going in circles here. just remember to AF on your subject
2-3 times and you should be golden! if you deliberately force your
camera's AF to "check and recheck" its AF results you won't have to
worry about this!

The autofocus uses a different and more complicated optical path which uses extra mirrors and lenses. If this path is not calibrated correctly the camera will systematically front or back focus. For the autofocus everythings looks perfect but unfortunately the image sensor is not in the spot the focus sensor expects if the camera is poorly calibrated.

WilbaW · Apr 23, 2009

Andreas Helke said:
WilbaW said:

Andreas Helke said:

The autofocus system is a mix of open loop and closed loop system.
The first step is a closed loop operation mode to look for the
aproximate focus. If the lens is already reasonably close to correct
focus this step is skipped.

Click to expand...

So you would say that the experiment in the OP only relates to that
step?

Click to expand...

Ir may well be. I could not repeat the experiment because my lenses
all focus too fast.

Andreas Helke said:

Andreas Helke said:

The final focus measurement is open loop. The camera measures the
distance. Very precise is the equipment is calibrated correctly and
you have a high contrast edge to focus on but often seriously wrong
for low contrast subjects. After that open loop distance measurement
step it uses a closed loop again for the actual focusing.

Click to expand...

I often see this stated - is there a definitive source?

Click to expand...

The information comes from Canon technical literature.

Any references?

Unfortunately there are some significant translation problems with some Canon literature. I think it would be wise to be skeptical about things described as "open-loop" in Canon documents.

Andreas Helke said:
From ovserving how my camera behaves and misbehaves I believe the
description is correct.

Fair enough, but ... when I do the experiment, I find that during what seems to be the final "actual focussing" step, the process can be interupted by a change of subject. I guess you could argue that if that's happening like an interupt (as in computer programming), rather than strictly as a result of a test in a control loop, then it's not closed-loop.

What misbehaviour do you see?

Andreas Helke said:
I called the lens adjustment closed loop because it relies on
feedback from measurements in the lens itself. But its just done once
and not until the system stabilizes somewhere. So its not a normal
feedback driven closed looped operation. This operation explains why
you can often notice an extra lens movement if you have a micromotor
lens.

I'm not convinced that that's the only or best explanation for that behaviour. I'd really like to see the Canon document(s) it comes from.

WilbaW · Apr 24, 2009

WilbaW said:
Erik Magnuson said:

WilbaW said:

I understand the argument, but without knowing a lot more about what
lens calibration actually does ...

Click to expand...

Depends on the lens and the issue. For modern lenses, it's
re-programming some tables in the lens EEPROM. For others it may be
physical shims or soldering bridges over jumper points (as someone
found in the the 50mm f/1.8 Mk I service manual).

Click to expand...

That's how you do it, not what it does.

The question is, what do those methods of adjustment achieve - what
aspects of the lens's performance are being altered and in what way?

That service manual is interesting ( http://www.ozerki.net/photosight/PhotoDocs/Manuals/partlist/Lens/CanonServiceManual (EF-50-28-15).zip).

These paragraphs give us something -

"In the EOS system, the difference between the AF focus and the optical best focus has been determined for each lens type and the information written into the lenses Read Only Memory (ROM) so that correction for the difference at maximum aperture is made electronically.

"In actuality, in addition to this type difference, there is a difference between individual lenses within each type, which can be noticeable if not corrected, At the factory, correction is written into the individual lens' ROM with a expensive, special tool. This is called the 'Best Focus Adjustment'. Because of the tooling cost involved, this adjustment will not be a part of the service procedure. In its stead, ... [two switches are set]." (EF50/1.8, VI-7, including original grammatical errors.)

So I get from that that there is a known lens-type difference, with a per-unit variation, between the AF focus and the optical best focus. That difference depends on aperture, and is corrected within each lens according to its own settings. This strongly suggests that it's invalid to infer that the system is open-loop because of any focus errors.

Erik Magnuson · Apr 24, 2009

WilbaW said:
So I get from that that there is a known lens-type difference, with a
per-unit variation, between the AF focus and the optical best focus.
That difference depends on aperture, and is corrected within each
lens according to its own settings. This strongly suggests that it's
invalid to infer that the system is open-loop because of any focus
errors.

Your conclusion does not follow from those statements. If a lens needs to be (re) calibrated, it's because those per lens adjustments are not correct for that particular lens. Don't get so hung up on the terminology "open loop" vs. "closed." As previously stated, the real system is not that simple.

The important information is that the system depends on the lens moving to where it's told and the system assumes that if this happens, correct focus will be achieved. So it does not typically verify the focus solution again. If the lens did not move the correct amount (because it's internal table is wrong), you will have FF or BF. A system that does check the solution after each movement (e.g. contrast detect) will not have the same issue.

--
Erik

WilbaW · Apr 25, 2009

Erik Magnuson said:
WilbaW said:

So I get from that that there is a known lens-type difference, with a
per-unit variation, between the AF focus and the optical best focus.
That difference depends on aperture, and is corrected within each
lens according to its own settings. This strongly suggests that it's
invalid to infer that the system is open-loop because of any focus
errors.

Click to expand...

Your conclusion does not follow from those statements.

Perhaps it needs a few more words to remove any ambiguity. Try, "That difference depends on the exposure aperture" (not the lens max. aperture).

Anyway, please explain how a different conclusion is more valid. (Don't forget that what were discussing at this point is the idea that "FF and BF still happen" because the AF system is fundamentally open-loop. My previous post explains how focus errors can still occur with a closed-loop system, either using a perfectly calibrated proxy sensor or using the image sensor, because the error is introduced at exposure time.)

Member said:
If a lens needs to be (re) calibrated, it's because those per lens adjustments
are not correct for that particular lens.

Exactly. And that explains how even though the phase error can be zero on a perfectly calibrated AF sensor, the image can still be out of focus with a particular lens, since the focus can shift during the exposure.

Member said:
Don't get so hung up on the terminology "open loop" vs. "closed." As
previously stated, the real system is not that simple.

It is a simple question - does the system use feedback from the sensor? The evidence and reason make a strong case for that being true. It does contradict a lot of conventional wisdom (e.g. the famous RDKirk post at Fred Miranda), which follows some questionable assumptions and is at times erroneous.

Member said:
The important information is that the system depends on the lens
moving to where it's told and the system assumes that if this
happens, correct focus will be achieved. So it does not typically
verify the focus solution again.

Do you have any direct firsthand evidence for this? Can it be tested?

Member said:
If the lens did not move the correct amount (because it's internal table is
wrong), you will have FF or BF.

If that were true, iterations would improve the focus. My tests show that that doesn't work.

Member said:
A system that does check the solution after each movement (e.g. contrast
detect) will not have the same issue.

The quotes from the manual suggest quite clearly that a CD system would have the same problem, since the error is introduced at exposure time, via the aperture stop-down - aperture-related focus shift.

Erik Magnuson · Apr 25, 2009

WilbaW said:
Perhaps it needs a few more words to remove any ambiguity. Try, "That
difference depends on the exposure aperture" (not the lens max.
aperture).

All that means is that some lenses have significant focus shift where the optimal focus wide-open (where the AF works) is not the same as the optimal focus stopped down. So taking aperture is one factor in the internal movement tables.

WilbaW said:
Anyway, please explain how a different conclusion is more valid.

Focus shift is not the most common reason for FF or BF. In those cases where it is a factor, it's also a problem for good manual focus.

WilbaW said:
(Don't forget that what were discussing at this point is the idea
that "FF and BF still happen" because the AF system is
fundamentally open-loop. My previous post explains how focus errors
can still occur with a closed-loop system, either using a perfectly
calibrated proxy sensor or using the image sensor, because the error
is introduced at exposure time.)

But most of the complaints we've seen posted here are for cases where the user is shooting wide open where the taking and shooting aperture are the same. So focus shift cannot be the problem we're discussing.

WilbaW said:
Exactly. And that explains how even though the phase error can be
zero on a perfectly calibrated AF sensor, the image can still be out
of focus with a particular lens, since the focus can shift during the
exposure.

Yes, it can be for a few fast lenses slightly stopped down where the amount of focus shift is potentially larger than the DOF. But it absolutely would not apply to any focus errors when shooting wide open. BTW, contrast detect AF is also dome with the lens wide open.

WilbaW said:
It is a simple question - does the system use feedback from the
sensor?

But it's not a simple answer ;-)

WilbaW said:
The evidence and reason make a strong case for that being
true. It does contradict a lot of conventional wisdom (e.g. the
famous RDKirk post at Fred Miranda), which follows some questionable
assumptions and is at times erroneous.

It's an oversimplification, but one that is useful for modeling the AF system for practical use.

WilbaW said:
Do you have any direct firsthand evidence for this? Can it be tested?

Well, first one needs a defective lens. Alas, I don't have any.

WilbaW said:
If that were true, iterations would improve the focus. My tests show
that that doesn't work.

That depends on exactly what the camera-lens communication consists of. For example, if the positioning is absolute, the camera would see that the lens is already at the target position (or the lens would immediately return the "I'm there" signal) and no lens movement would happen. It would just confirm the "wrong" focus.

WilbaW said:
The quotes from the manual suggest quite clearly that a CD system
would have the same problem, since the error is introduced at
exposure time, via the aperture stop-down - aperture-related focus
shift.

Well, yes, if that were the most common cause of FF or BF.

--
Erik

WilbaW · Apr 25, 2009

Erik Magnuson said:
WilbaW said:

Perhaps it needs a few more words to remove any ambiguity. Try,
"That difference depends on the exposure aperture" (not the lens
max. aperture).

Click to expand...

All that means is that some lenses have significant focus shift where
the optimal focus wide-open (where the AF works) is not the same as
the optimal focus stopped down. So taking aperture is one factor in
the internal movement tables.

Exactly. So if we have a perfectly calibrated body, we can still get a focus error with a closed-loop system (i.e. one in which the phase difference is checked during the focus motion until it is within acceptable limits). Concluding that only an open-loop system can exhibit FF and BF errors is invalid. That's the point of this part of the discussion.

Member said:
Member said:

It is a simple question - does the system use feedback from the
sensor?

Click to expand...

But it's not a simple answer ;-)

Sure it is, the answer is yes or no. A simple experiment (provided at the beginning of the thread) proves that it's yes.

If you want to go inside the system and consider the AF sensor, the image sensor, the lens, and the control algorithms individually, then it certainly gets complicated, but taking all of those components together as a system , you can simply observe its black-box behaviour.

Member said:
Member said:

If that were true [FF and BF caused by uncorrected lens motion], iterations
would improve the focus. My tests show that that doesn't work.

Click to expand...

That depends on exactly what the camera-lens communication consists
of. For example, if the positioning is absolute, the camera would see
that the lens is already at the target position (or the lens would
immediately return the "I'm there" signal) and no lens movement would
happen. It would just confirm the "wrong" focus.

That doesn't make sense to me. If you start the focus process all over again (that's the iteration we're talking about), a fresh measurement is taken from the sensor, and the system will act on that, not on where the lens is or thinks it is.

Erik Magnuson · Apr 25, 2009

WilbaW said:
Exactly. So if we have a perfectly calibrated body, we can still get
a focus error with a closed-loop system (i.e. one in which the phase
difference is checked during the focus motion until it is within
acceptable limits).

In certain rare cases with certain lenses and aperture combinations - but never with the lens wide open which is where the majority of problems are seen.

WilbaW said:
Concluding that only an open-loop system can
exhibit FF and BF errors is invalid. That's the point of this part of
the discussion.

OK, how about if we say that only an open-loop system would exhibit FF or BF errors when the taking aperture is the same as the focusing aperture?

WilbaW said:
Sure it is, the answer is yes or no. A simple experiment (provided at
the beginning of the thread) proves that it's yes.

Yes or no is not an interesting answer -- except to be pedantic about classifying the system with a particular binary label.

WilbaW said:
That doesn't make sense to me. If you start the focus process all
over again (that's the iteration we're talking about), a fresh
measurement is taken from the sensor, and the system will act on
that, not on where the lens is or thinks it is.

What is the "action" that is taken? It's got to be a lens movement command. But what if the lens thinks its already at the correct position? Will it move?

--
Erik

WilbaW · Apr 25, 2009

Erik Magnuson said:
WilbaW said:

So if we have a perfectly calibrated body, we can still get a focus
error with a closed-loop system (i.e. one in which the phase
difference is checked during the focus motion until it is within
acceptable limits).

Click to expand...

In certain rare cases with certain lenses and aperture combinations -
but never with the lens wide open which is where the majority of
problems are seen.

Correct - aperture-related focus shift should not appear with an exposure at maximum aperture. "Rare" and "majority" I'm not going to enter into.

Member said:
Member said:

Concluding that only an open-loop system can exhibit FF and BF errors
is invalid. That's the point of this part of the discussion.

Click to expand...

OK, how about if we say that only an open-loop system would exhibit
FF or BF errors when the taking aperture is the same as the focusing
aperture?

That follows from what we know, yes.

Member said:
Member said:

Sure it is, the answer is yes or no. A simple experiment (provided at
the beginning of the thread) proves that it's yes.

Click to expand...

Yes or no is not an interesting answer -- except to be pedantic about
classifying the system with a particular binary label.

It's far from trivial since many interpretations and assertions about what's happening are based on the answer. If everything else that follows is based on the wrong answer, then the interpretations are also probably wrong.

What is your answer to this question - how does your model of the black-box system account for the behaviour observed in the experiment? (During focussing under normal conditions, the system responds to a change in subject distance by focussing at the new subject distance.) This is the REALLY interesting question.

Member said:
Member said:

That doesn't make sense to me. If you start the focus process all
over again (that's the iteration we're talking about), a fresh
measurement is taken from the sensor, and the system will act on
that, not on where the lens is or thinks it is.

Click to expand...

What is the "action" that is taken? It's got to be a lens movement
command.

If required, yes.

Member said:
But what if the lens thinks its already at the correct position? Will it move?

Why would the lens be asked what it thinks? Surely the first step is to assess the focus error as perceived by the AF sensor?

Erik Magnuson · Apr 26, 2009

WilbaW said:
Correct - aperture-related focus shift should not appear with an
exposure at maximum aperture. "Rare" and "majority" I'm not going to
enter into.

Lens reviews will mention this because it's a problem for lens testing:

focus shift will affect MF and contrast-detect AF in the same manner. You can see how frequently it's referenced in that context to get an idea of it's rarity.

WilbaW said:
It's far from trivial since many interpretations and assertions about
what's happening are based on the answer. If everything else that
follows is based on the wrong answer, then the interpretations are
also probably wrong.

No, it's more like the difference between Newton's and Einstein's models of physics. Newtonian physics works fine for most everyday applications on earth even though they are not strictly correct.

WilbaW said:
What is your answer to this question - how does your model of the
black-box system account for the behaviour observed in the
experiment?

Simple: it responds to feedback only when the target solution changes during lens movement. But for a static subject it still only computes the AF solution once and if that solution is wrong, you will get misfocus. It does not recompute a new solution unless the input changes.

WilbaW said:
Why would the lens be asked what it thinks? Surely the first step is
to assess the focus error as perceived by the AF sensor?

We're past this step. What's the 2nd step? The AF system tells the lens to move somewhere.

--
Erik

WilbaW · Apr 26, 2009

Erik Magnuson said:
WilbaW said:

What is your answer to this question - how does your model of the
black-box system account for the behaviour observed in the
experiment?

Click to expand...

Simple: it responds to feedback only when the target solution changes
during lens movement.

"It responds to feedback" means the system is closed-loop, by definition. Simple.

Member said:
But for a static subject it still only computes the AF solution once and if that
solution is wrong, you will get misfocus. It does not recompute a new solution
unless the input changes.

Can you think of an experiment that can show this to be true? Or if you want to really get scientific about it, can you think of an experiment which would show it to be false?

I seriously doubt that you can, since it seems to me to be an assertion based speculation rather than on observation of unambiguous behaviour. A system which repeatedly checks the focus error and revises its target focus distance would behave identically in that situation to the system you believe in. Except that it would accomodate the subject-shift inherently as part of its normal operation, rather than relying on a monitor and interupt kind of control algorithm.

And then we come to AI Servo - how is the operation of AI Servo different to One Shot in your model?

Member said:
We're past this step. What's the 2nd step? The AF system tells the
lens to move somewhere.

If it needs to move, yep.

Erik Magnuson · Apr 26, 2009

WilbaW said:
"It responds to feedback" means the system is closed-loop, by
definition. Simple.

It doesn't respond to feedback - there are certain conditions that cause it to abort the current operation and start a new one. Only a single lens movement command is given for any given AF solution.

WilbaW said:
Can you think of an experiment that can show this to be true?

It's the simplest explanation for FF or BF wide open.

WilbaW said:
A system which repeatedly checks the focus error and
revises its target focus distance would behave identically in that
situation to the system you believe in.

No, because it would see the FF or BF error before stopping the lens and continue to adjust.

WilbaW said:
And then we come to AI Servo - how is the operation of AI Servo
different to One Shot in your model?

Continuous AF is just a continuous sequence of discrete look/compute/move operations. (I'm sure it's more complicated then that, particularly with predictive subject tracking.)

WilbaW said:
If it needs to move, yep.

Exactly. If the lens thinks it's already at the commanded position, will it move?

--
Erik

WilbaW · Apr 27, 2009

Erik Magnuson said:
WilbaW said:

"It responds to feedback" means the system is closed-loop, by
definition. Simple.

Click to expand...

It doesn't respond to feedback ...

Hey, come on, you can't say "it responds to feedback" one minute, and "it doesn't respond to feedback" the next, just because you don't like the consequences.

Member said:
... - there are certain conditions that cause it to abort the current operation
and start a new one.

Yep, it responds to feedback (or would you prefer input?), from the AF sensor.

Member said:
Only a single lens movement command is given for any given AF solution.

I understand the theory, but it's not the only one that works, nor is it the most elegant, practical, or reasonable.

Member said:
Member said:

Can you think of an experiment that can show this to be true?

Click to expand...

It's the simplest explanation for FF or BF wide open.

Except for other factors, such as softness wide open (e.g. EF 50/1.8), spherical aberation, and who knows what, that can make it difficult to establish an accurate focus by phase difference. Anyway, there's no way to verify the "one command" theory without getting right inside the gubbins, and no definitive statement from Canon, so it's a belief.

Member said:
Member said:

A system which repeatedly checks the focus error and
revises its target focus distance would behave identically in that
situation to the system you believe in.

Click to expand...

No, because it would see the FF or BF error before stopping the lens
and continue to adjust.

Only in a perfect world in which your system was imperfect (not able to calculate the required lens position correctly).

Let's leave aside the FF/BF things for a while because it's not getting us anywhere.

Here's a summary of where we're at (I think). In your model of a static focus situation, the sensor says the phase difference is A, the AF system works out that the lens has to be at X, and tells the lens to move there. When the lens says it is at X, the system confirms focus.

In my model, the sensor says the phase difference is A, the AF system works out that the lens has to be at X, and tells the lens to move there. While the lens is moving to X, the sensor says the phase difference is B, the AF system works out that the lens has to be at X (same as before), and tells the lens to move there, and so on. When the phase difference is zero, the system confirms focus.

The observer can't tell the difference between the two systems, because they both move the lens to the correct position to achieve focus.

In a non-static focus situation (as in the experiment), while the lens is moving to X, in your model the sensor says the phase difference is B, the AF system works out that the lens has to be at Y, and tells the lens to move there .... So does mine, exactly the same as in the static situation.

Both models are closed-loop by definition. The only real differences are that you see an in-process change of subject distance as an exception, and focus confirmation as a lens thing rather than an AF sensor thing.

Just checking - you know that closed-loop doesn't just mean that a series of instructions are executed in a loop, right? That's about implementation. The loop in closed-loop is about whether the system responds to changes in the environment during operation.

For your system to know that the subject distance has changed, there is probably a test implemented in a loop somewhere, so even if you follow that misconception of closed-loop, your system would be closed-loop anyway.

Member said:
Member said:

And then we come to AI Servo - how is the operation of AI Servo
different to One Shot in your model?

Click to expand...

Continuous AF is just a continuous sequence of discrete
look/compute/move operations.

OK, the only modifications necessary to the two models is that we just don't confirm focus.

Do you agree that your black-box model in AI Servo mode is closed-loop? If not , why not?

Member said:
(I'm sure it's more complicated then that,
particularly with predictive subject tracking.)

Sure. Essentially that just requires using information from a sequence of readings from the sensor.

Member said:
Member said:

If it needs to move, yep.

Click to expand...

Exactly. If the lens thinks it's already at the commanded position,
will it move?

We've been going around in circles so long with this bit and I still have no idea what you're thinking or trying to get at. How about you just come out and say it?

Erik Magnuson · Apr 28, 2009

WilbaW said:
Hey, come on, you can't say "it responds to feedback" one minute, and
"it doesn't respond to feedback" the next, just because you don't
like the consequences.

Sure I can. It depends on how you use the word "feedback" and where you draw the lines in the system.

WilbaW said:
Yep, it responds to feedback (or would you prefer input?), from the
AF sensor.

But the input is not used to directly control the lens movement, e.g. tell the lens how fast to move or interactively when to stop. That's the feedback that is typically considered to be the hallmark of a closed-loop system.

WilbaW said:
I understand the theory, but it's not the only one that works, nor is
it the most elegant, practical, or reasonable.

That would depend on the implementation constraints, wouldn't it?

WilbaW said:
Except for other factors, such as softness wide open (e.g. EF
50/1.8), spherical aberation, and who knows what, that can make it
difficult to establish an accurate focus by phase difference.

You need to explain why these factors affect only phase AF and not contrast detect AF. Softness, spherical aberration and who know what (focus shift), would affect both systems identically.

WilbaW said:
no definitive statement from Canon, so it's a belief.

Yes, it's a theory. It may be wrong or incomplete, but it fits the observed behavior and what little info from Canon we have.

WilbaW said:
Let's leave aside the FF/BF things for a while because it's not
getting us anywhere.

You mean you have not accounted for it in your theory ;-)

WilbaW said:
Just checking - you know that closed-loop doesn't just mean that a
series of instructions are executed in a loop, right? That's about
implementation. The loop in closed-loop is about whether the system
responds to changes in the environment during operation.

As I've said, I'm not worried about the precise labeling of the system. The difference between our theories is how tightly coupled (iterative) the lens movement and the AF system are. For phase detect, I say they are very loosely coupled (e.g. one calculation, one move command, typically no reconfirm); for contrast detect they are more tightly coupled (move, check, move, check, etc.)

WilbaW said:
your system would be closed-loop anyway.

If it makes you happy. This "is so" vs. "is not" is getting tedious.

WilbaW said:
We've been going around in circles so long with this bit and I still
have no idea what you're thinking or trying to get at. How about you
just come out and say it?

I've said it several times. We're just not communicating.

--
Erik

WilbaW · Apr 29, 2009

Erik Magnuson said:
WilbaW said:

Let's leave aside the FF/BF things for a while because it's not
getting us anywhere.

Click to expand...

You mean you have not accounted for it in your theory ;-)

I mean you haven't explained your model well enough for me to understand where you're going wrong. :-D

Try this. Two perfect AF systems with perfect sensors and optically imperfect but well calibrated lenses.

Your system:

1. Measures the phase difference and calculates the lens position to make it zero.
2. Tells the lens to go there.

3. Monitors the phase difference, and restarts the process if a significant change in subject distance is detected.

4. The lens reaches the position that eliminates the phase difference at the AF sensor.

5. Confirms focus, and gets the lens to make any final correction based on it's own calibration.
6. Any remaining focus imperfections are due to other factors in the lens.

My system:

1. Measures the phase difference and calculates the lens position to make it zero.
2. Tells the lens to go there.
3. Rechecks the phase difference and required lens position.

4. The lens reaches the position that eliminates the phase difference at the AF sensor.

5. Confirms focus, and gets the lens to make any final correction based on it's own calibration.
6. Any remaining focus imperfections are due to other factors in the lens.

Q1. What is the difference between the end-states of the two systems?

Q2. What is the difference between the observable behaviour of the two systems?

Erik Magnuson · Apr 29, 2009

WilbaW said:
I mean you haven't explained your model well enough for me to
understand where you're going wrong. :-D

It's to account for most known cases of FF and BF that is one of the key rationales for the pseudo "open loop" theory.

WilbaW said:
Try this. Two perfect AF systems with perfect sensors and optically
imperfect but well calibrated lenses.

Try this: the same AF system, and two arbitrary lenses of the same model. Shooting wide open, one FFs a little bit one BFs a lot. Using contrast detect, both lenses focus identically. Sending the BFing lens in for calibration brings it into tolerance. Since the optical properties of the lenses are identical, how could they ever work differently in the first place and what did they adjust to fix the second lens? According to a Canon service tech, the calibration involved no mechanical adjustments.

WilbaW said:
Your system:
5. Confirms focus, and gets the lens to make any final correction
based on it's own calibration.

These steps will be the other way 'round. And focus confirm is really just the signal that the lens (thinks it has) has reached the commanded position. The AF solution is not typically rechecked (there are exceptions).

WilbaW said:
Q1. What is the difference between the end-states of the two systems?

See my hypothetical above. What are the "other factors" you are proposing about the lens that could account for these conditions? Note that since the lenses are optically identical (aberrations are identical) and shot wide open (focus shift) none of the previous causes you have mentioned apply.

WilbaW said:
Q2. What is the difference between the observable behavior of the
two systems?

If the solution is always rechecked, significant FF or BF wide open could never happen because the solution would be incorrect and the lens would be commanded to move again. If the lens was defective in it's ability to move accurately, this would cause the lens to "hunt" between two incorrect positions and the 2nd check would also likely fail.

--
Erik

brightcolours · Apr 29, 2009

Erik always just goes on and on, blindly repeating, not thinking.

One thing that gets said over and over is the following:

That the camera checks the phase difference and "calculates" what the lens has to do.

It does not, and it can not.

Phase detection can see a "value" difference to focus, including the direction. But, what it can not do, is know what that value means for the lens in question.
The lens in question ALSO can not know what the value means.

Simple experiment to illustrate this:

With a 12mm extension tube, my 70-200 f4 L USM lens thinks that 2.5 meters away is 10 meters away.
It thinks that 1 meter away is 1.2 meters away.

the extension tube allows me to focus about 20 to 30 cm closer, but I lose infinity focus BIG time.

The normal 1.2-10 meter focus path gets used/compressed for the 1-2.5 meter focus path.

However, the field of view, and what the AF sensors see in respect to phase difference, almost does not change.

So... the phase difference is the same, what the lens has to move changes a LOT.

And yet, the AF speed of the lens stays the same. And both the lens and the camera do not know of the existence of the mounted extension tube.

This example shows that the lens, nor the camera, actually know how to interpret the phase difference.

How does it work then?
Quite simple, and quite effective (have written this many times before).

The camera tells the lens to go through steps moving the lens, to get to a focus area where the camera can see edges/contrast, so it can determine phase difference.
All the while, the camera checks phase difference, till it can see.

Now, when it can detect a phase difference, the camera also can see the direction. It now tells the lens to do a step into that direction. It detects another phase difference. From those two readings, it calculates the percentage of last lens movement needed to reach focus (or as you call it, "zero").

Now the camera will tell the lens to move that percentage, and focus lock is declared.

Now why does the camera not check focus again? Speed reasons. If AF would not be reached, the camera would have to correct again. And after that, again. With well working and well calibrated lenses, the results are a good compromise between focus and speed.

And why do some lenses hunt? Because the steps it makes to reach a state where the camera can actually detect edges gets overstepped, and then you have a hunting lens. Hunting occurs sooner with lower light conditions (less contrast), and with lenses that have motor systems that make bigger steps (like my Tamron 90mm f2.8 macro).

There are MANY other factors that show that phase detection AF in DSLRs is not a one step, one movement affair, but listing them all is pointless with Erik. He never really likes to react really to issues raised.

MM1 · Apr 29, 2009

brightcolours said:
Now why does the camera not check focus again? Speed reasons. If AF
would not be reached, the camera would have to correct again. And
after that, again. With well working and well calibrated lenses, the
results are a good compromise between focus and speed.

This would be so easy to implement in firmware: finite number of rechecks so that everyone would balance his need of speed and accuracy.

Sometimes, I would welcome slower but more accurate AF more, than lightnitng fast one, which produces something that can be called "a rough estimate."

--
Cheers,
Martin

Open-loop or Closed-loop Phase Detect Autofocus?

Forum Pro

Senior Member

Senior Member

Forum Pro

Forum Pro

Erik Magnuson

Guest

Forum Pro

Erik Magnuson

Guest

Forum Pro

Erik Magnuson

Guest

Forum Pro

Erik Magnuson

Guest

Forum Pro

Erik Magnuson

Guest

Forum Pro

Erik Magnuson

Guest

Forum Pro

Erik Magnuson

Guest

Forum Pro

Senior Member

Keyboard shortcuts