Busted! The Myth of Open-loop Phase-detection Autofocus

Introduction

When people ask for help with focus problems in the Canon forums here, it is often stated that phase-detection (PD) autofocus (AF) operating in a mode that locks focus before enabling the shutter to fire (e.g. "One-Shot" for Canon, "AF-S" for Nikon, Pentax, and Sony, "S-AF" for Olympus, "S" for Sigma), works like this –

The Open-loop Model

  1. The AF sensor takes a look at the subject,
  2. the AF system calculates exactly how far the focus has to change to bring the subject into focus,
  3. the AF system instructs the lens to make that change,
  4. the lens tells the AF system that it has completed the movement,
  5. the AF system confirms focus*, with no second look at the subject.

* Focus is typically confirmed by the camera flashing the active AF point, turning on the green focus confirmation light in the viewfinder, beeping, and enabling shutter firing if it was disabled.

This belief, at least in the Canon world, seems to be based on the famous RDKirk posts on the Fred Miranda forums, and many deliberately non-committal or misleading statements from Canon guru Chuck Westfall. Chuck has stated that Canon will not disclose how One-Shot works, although elsewhere he describes closed-loop operation with AF-Assist, and that "focus confirmation is based on successful completion of the lens drive command". Chuck has been invited to comment through various online forums where he used to participate, but the official word from Canon is that, "Mr Westfall does not currently have a blog and does not contribute to any open web forums." He has also not replied to direct e-mail. There doesn't seem to be any confusion or doubt in the Nikon world, since it's obvious to most that a system with the focus motor in the body must be closed-loop.

The "one measurement, one movement" model is superficially plausible, and descriptions of the rangefinder principle give the impression that it is that simple. But is it a "flat Earth" model, seeming to work well on a limited scale, but failing to predict and explain the broader reality? How can we test it?

In control theory, a "one measurement, one movement" design is classified as an open-loop system, meaning that feedback about the actual state of the thing being controlled isn't used to refine what the system is doing. It's like firing a cannon – you aim it with all possible care and skill, but once the cannonball has left the muzzle you can't control where it's going to land.

In a continuously focussing mode (e.g. "AI Servo" for Canon, "AF-C" for Nikon, Pentax, and Sony, "C-AF" for Olympus, "C" for Sigma), there's no question that the AF system makes use of continuous feedback from the AF sensor to track the subject – that's definitively a closed-loop system (see figure 3 at Nikon's description of their Overlap Servo Predictive Focus Tracking System for a good depiction of a sophisticated closed-loop control system). It's like firing a guided missile – it doesn't matter so much where you aim because the missile knows where it has to go and how to get itself there.

Scientific American were in no doubt that all forms of PD AF are a closed-loop system – "A logic circuit constantly monitors [the distance between the images on the AF sensor elements] and drives a motor that spins the focusing ring of the lens, shifting the focus. When the separation hits a predetermined value, the logic circuit stops the motor and flashes lights in the viewfinder to indicate that the image is focused." (Focusing in a Flash, Scientific American Magazine, August 2000.) Canon Europe acknowledge this on their Infobank page which describes How AF works – "If there is no deviation between the two images seen by the sensors, the lens is focused."

From what I can gather from various sources, Honeywell patented PD AF in the 1970s, sued Minolta for patent violation and won (1987-91, after the 1985 release of the Maxxum/Dynax/Alpha 7000), and presumably the other manufacturers signed up to pay royalties for the system rather than face the same battle. Testing of DSLRs from Canon, Nikon, Olympus, Pentax, Sigma, and Sony (Minolta) shows they do all work fundamentally the same (see Verified Bodies). I've read some of those Honeywell patents but I wasn't able to get a clear idea of what was intended on this question of open versus closed-loop.

How would an ideal closed-loop PD AF system work in a focus-locking mode?

The Closed-loop Model

  1. The AF sensor takes a look at the subject,
  2. the AF system calculates how far the focus has to change to bring the subject into focus,
  3. the AF system instructs the lens to make that change,
  4. the AF system continuously* repeats steps 1 to 3 until...
  5. the AF system confirms focus** when the AF sensor sees that the subject is in focus.

* The AF system does not wait for the lens to finish a focus movement before taking another look, it looks, calculates, and instructs "all the time" during focus movement.

** Notice how, if you leave out step 5, and change step 2 to predict where a moving subject will be when the shutter opens, you have a perfect model of a continuous-focus mode.

If you're uncomfortable with the terms "open-loop" and "closed-loop", think "one measurement, one movement, no second look", and "continuous measurement, continuous movement, final look", instead.

When I first came across the assertion that focus-locking modes were open-loop I thought it sounded implausible –

  • Why would the engineers use an open-loop system when they already have a closed-loop system doing the exact same job? It makes no sense whatsoever to not check the result of a dynamic action requiring precision, when you have the ability to do so quickly and efficiently. (This is sufficient counter-argument all by itself, but since we're dealing with faith we need to challenge it with empirical proof.)
  • How is it possible for the AF system to have enough accurate information to exactly determine the required focus, for every combination of nominal focal length, current focus position, subject distance, and exposure aperture?
  • How could the AF system possibly calculate exactly where focus needs to be when the optical situation is radically different from normal, such as when using an extension tube, close-up attachment, or teleconverter?
  • How would such a system account for wear and backlash in the focus drive mechanism, particularly if the lens is driven by a motor in the body?

But since it seemed everyone was adamant that it was true, I didn't think much more about it. Then I read about a photographer trying to take a shot of a squirrel with a Canon DSLR. He had the squirrel in his sights and half-pressed the shutter button, but before the lens made it to where the squirrel was the squirrel jumped out of sight.

Where should the focus have ended up? If it were an open-loop system, the focus destination was set when the shutter button was half-pressed, so focus would end up where the squirrel was at that time, and focus would be confirmed when it reached that distance. If it were a closed-loop system, the focus destination would depend only on what the AF sensor could see, so it would end up on whatever was under the active AF point, in this case, the background.

The squirrel photographer found that focus was confirmed on the background, suggesting that it is a closed-loop system.

Scope, Definitions, and Assumptions

The purpose of this article is to give you the means to gather your own evidence, to reach your own conclusions about whether focus confirmation depends on what the AF sensor sees (and therefore whether the AF control system is open or closed-loop), so I'm not going to spell out the results you should see from each test. The bonus questions are there to give you other ways to enrich your understanding of the evidence, and I'm not going to spoon-feed you the results of them either. However, at the end of each test I will provide a summary of what it tells us with respect to the competing hypotheses.

I'm not trying to explain why a closed-loop system requires lens calibration for accuracy. For the purpose of this article, I accept the fact that lens calibration is essential for accurate focus, whether the system is open or closed-loop.

The instructions in the tests assume that the camera is set up to autofocus on the shutter button. You can substitute "press the AF-ON (or '*') button" for "half-press the shutter button" in test instructions.

When I refer to "the subject", I'm talking about general photographic situations, whereas when I refer to "the target", I'm talking about specifically what we see when we do one of the tests in controlled conditions.

When I refer to "the AF system", I mean the embodied logic that interprets the signal from the AF sensor, the current settings, and the photographer's actions, and decides what to do with all that, including commanding the lens to move the focus.

I'm not modelling or testing the body-to-lens communication protocol, what happens inside the lens to comply with a command from the AF system, or whether there are any feedback loops involved in any aspect of those things. After reading the article and doing the tests you should be able to work out why those things don't matter to the question of whether focus is confirmed via feedback from the AF sensor.

I'm not modelling or testing the predictive element of continuous-focus modes.

All tests are conducted under "good" conditions, with well-illuminated high-contrast and detailed targets, which present no challenge to the normal operation of the AF system. The tests are not aimed at finding out what happens outside of those conditions, although doing so can be interesting. Please let me know if you discover something.

The Squirrel Test

It's easy to test for open-loop behaviour for yourself. All you have to do is make predictions according to the competing hypotheses and execute an experiment based on the scenario to see which prediction comes true.

Scenario – the target changes position between the beginning and end of focus movement.

HypothesisPrediction
Open-loop Focus is confirmed when it reaches where the target was when AF was triggered.
Closed-loop Focus follows the target and is confirmed when the AF sensor sees that the target is in acceptable focus.

Method

  1. Set up a target near the minimum focus distance (MFD) in front of another in the far distance (I use a marked Post-It note on my window pane with my neighbour's tree behind it).
  2. Use a long slowish-focussing lens, like the Canon EF-S 55-250 @ 250mm or the Nikon AF-S 55-200 @ 200mm.
  3. Set the camera to a focus-locking mode (One-shot or AF-S) using a single AF point.
  4. Focus on the distant target.
  5. Aim at the close target.
  6. Half-press the shutter button and as soon as focus starts to move, swing the AF point directly onto the distant target*.
  7. Try it again continually swinging the AF point from target to target before it gets into focus.

* Be aware that the scope of an AF point is generally much larger than the corresponding box in the viewfinder, so you might have to swing further than you would think to give the AF point a clear view of the distant target.

Is focus confirmed when it reaches where the first target was, or does focus confirmation depend on what the AF sensor sees?

Bonus Questions

  1. Does the focus have to reach the close target before it heads back towards the distant target?
  2. If the focus does reach the close target before it heads back towards the distant target, is focus confirmed when it reaches there?
  3. Instead of a target near the MFD (step 1), try one a few metres away, manual focus at the MFD, half-press the shutter button with the AF point on the far target then swing onto the close target before focus passes it.
  4. Repeat the test in a continuous-focus mode (AI Servo or AF-C, step 3) – what differences do you see?
  5. Repeat the test with a full-press of the shutter button (step 6) – what differences do you see?

Interpretation of Results

  • Focus follows the target and is confirmed when the AF sensor sees the target in focus.
  • Focus doesn't have to reach a destination before the AF system responds to a change in target distance.
  • The behaviours of focus-locking and continuous-focus modes are indistinguishable under the test conditions.
  • With a full-press of the shutter button, the shutter won't fire if the AF sensor sees an out of focus subject.

To get a hands-on sense of what it means that "focus is confirmed when the AF sensor sees the target in focus", play with this Autofocus: Phase Detection simulation. As you move the "Change the len's position" slider you are performing the role of the AF system. You can take a guess at how far you have to move the slider, but you don't get focus confirmation just because you moved it that far. You have to keep trying until you get the slider in the right position to superpose the red and green peaks, and when you do you get "In Focus!" confirmation.

The "Open-loop With Continuous Feedback" Misconception

When presented with this evidence, proponents of the open-loop model often say that the AF system is able to detect that the subject moved, and so it knows to re-start the open-loop process. In other words, the AF system does make use of continuous feedback from the AF sensor, and is therefore by definition a closed-loop control system.

The "No Second Look" Myth

Having accepted the evidence of the Squirrel Test in a focus-locking mode, the open-loop proponents often say that what they really mean is that once the lens has reached the position required to achieve focus on a static subject, the AF system never takes another look via the AF sensor to confirm that the subject is in focus. Bonus question 2 of the Squirrel Test shows us that reaching a focus destination does not result in focus confirmation if the AF sensor sees an out-of-focus subject.

The Extension Tube Test

It's generally not possible to reliably judge whether a second look is taken at the speed of normal operation, but one way you can is with an extension tube and slow-focussing equipment. For instance, with an EOS 650 (the first EOS body, released in 1987), a 13mm auto extension tube, and an EF 35-70 1:3.5-4.5 kit lens of the same era, the individual steps in the multi-step autofocus process are clearly audible and visible.

The extension tube alters the relationship between a focus motion and the resulting change in focus, so that each motion results in a larger change than the camera required. When you trigger AF, focus passes through the subject and ends up on the other side, not so far out of focus, then it repeatedly reverses to converge on the subject, like:

zzzzzzzzzzzzzzzzt ->

<- zzzzzzzzt

zzzzt ->

<- zzt

zt ->

<- z

Focus is confirmed after the last movement. This test also proves beyond doubt that PD AF is not a "one measurement, one movement" process.

The Intermittent Illumination Test

There is a way to radically slow down the AF process and make apparent the individual steps in the AF process with any camera and lens.

Scenario – the AF sensor can't see the static target during focus movement.

HypothesisPrediction
Open-loop Focus is confirmed without the AF sensor taking a second look at the target.
Closed-loop Focus follows the target and is confirmed when the AF sensor sees that the target is in acceptable focus.

Method

  1. Somewhere you can make dark, set up a target near the MFD.
  2. Manually focus the lens at infinity.
  3. Set the camera to a focus-locking mode (One-shot or AF-S) using a single AF point, and turn off AF-assist.
  4. Turn off the lights.
  5. Half-press the shutter button and hold it down for the rest of the test.
  6. Briefly illuminate the target*.
  7. Count the number of flashes it takes to confirm focus.

* With a Canon you can fire an external flash to illuminate the target, but Nikons, Pentaxes, and Sonys need a lot longer duraton, for instance, switching a desk lamp on and off. (Canons use flash for AF assistance, so it makes sense that they would respond to any brief flash of light. Nikons and Pentaxs use lamps, so it makes sense that they only respond to steady illumination. Sonys use flash but won't respond to a manually fired external flash, so I guess something else is going on in that case.)

Is focus confirmed after a single look at the target by the AF sensor?

Bonus Questions

  1. How many flashes does it take to confirm focus if you immediately repeat steps 5 to 7 (with the target still in focus)?
  2. Can you get focus confirmation if you move the target out of focus before each flash (step 6)?
  3. If you stop after step 6, is the target in focus? (Check in Live View or switch to MF and take a shot. If it isn't, try focussing closer in step 2.)
  4. What difference does it make if you use a short extension tube, a close-up attachment, or a teleconverter? (In step 1, make sure the AF system is able to autofocus on the target in good light.)

Interpretation of Results

  • It might take just one look and one movement to achieve focus, but there is always at least a second look to confirm focus.
  • Focus won't be confirmed if a look following a movement sees an out-of-focus target.

The Open-loop Final Step Test

I worked on this with Doug Kerr (author of Principle of the Split Image Focusing Aid), using Canon bodies. He found that there is one specific circumstance in which the final step is a movement instead of a look – when the subject is just out of focus and the shutter button is full-pressed rather than half-pressed. This makes sense as an optimisation for maximum speed of operation when required (the full-press indicates the photographer's desire to fire the shot ASAP). Here's a test that shows it.

Method

  1. Somewhere you can make dark, set up a target.
  2. Set the camera to a focus-locking mode (One-shot or AF-S) using a single AF point, and turn off AF-assist.
  3. Autofocus on the target.
  4. Disable Focus Search*.
  5. Switch to manual focus mode (MF).
  6. Manually defocus** in tiny steps until there is no beep on a half-press (hold it down for a few seconds to make sure).
  7. Switch to autofocus mode (AF).
  8. Turn off the lights.
  9. Full-press the shutter button and hold it down for the rest of the test.
  10. Briefly illuminate the target.
  11. Count the number of flashes it takes to fire the shutter.

* For Canon, use the Lens drive when AF impossible custom function. As far as I know, this test is not possible with Nikon bodies since you can't disable focus search. If you don't disable focus search in step 4, as soon as you press the shutter button in step 9 the lens will rack through its range and you will lose your carefully prepared focus position.

** This can be done with the focus ring on a "Full Time Manual" (Canon) or "Manual Override" (Nikon) lens, or by changing the camera-to-target distance with something like a macro slide.

Bonus Questions

  1. What happens if you half-press instead of full-press in step 9?
  2. How far do you have to defocus in step 6 to make it take two flashes to get the shutter to fire?

Interpretation of Results

  • For a target that is just out of focus, a Canon system will execute a "one measurement, one movement" process if and only if the shutter button was full-pressed.

Playing Golf

A golfing analogy helps to understand how the AF system behaves when we don't let the AF sensor see anything during focus movement.

GolfFocus-locking PD AF
If the ball is already in the cup, you don't have to play another stroke. If the AF sensor sees that the subject is in acceptable focus, the focus doesn't have to change.
If the ball is on the lip of the cup, you don't have to line up a formal putt, you can just tap it and be confident that it will go in. If the AF sensor sees that the subject is just out of focus, and the shutter button is full-pressed, the focus can be adjusted the tiny amount required and no further check is necessary before firing the shutter.
If the ball is in a good position on the green, a good golfer can sink it in one putt, but there's no guarantee and it might need more putts to put it away. If the AF sensor sees that the lens is significantly out of focus, the focus will be changed in an attempt to achieve focus, but the result will always be checked to know whether another adjustment is required.
If the ball is on the tee, it will generally take a drive, maybe some fairway shots, and one or more putts, to sink it. If the AF sensor sees that the lens is grossly out of focus, the AF system will execute a series of increments converging on focus.

Short focal length lenses behave as if they are on the green no matter how far they are out of focus, and some only ever take a putt and a look to check that the ball is in the cup. Longer lenses behave as if they are on the tee, for instance, taking three to five distinct movements to focus from infinity to near minimum focus distance or vice versa.

Continuous Illumination

When autofocussing with normal levels of continuous illumination, the steps in the focus process are generally not apparent – most lenses appear to make one continuous movement towards focus. If you rest your finger on the focus ring sometimes you can feel distinct movements, and some lenses often show a twitch at the end. This perfectly fits with what you would predict using the open-loop hypothesis. A twitch is interpreted as the lens correcting any discrepancy between where it was told to go and where the first focus motion ended up. This view is supported by the myth that the notoriously unreliable and twitchy Canon EF 50mm f/1.8 II has a crude focus mechanism, which is believed to be unable to position the focus accurately or efficiently.

But this apparent single movement also perfectly fits with what you would predict using the closed-loop hypothesis with continuous feedback (focus instructions can be superseded before completion, as shown in the Squirrel Test). So the twitch is simply a visible final movement in a series of superseding instructions which converge on the AF sensor seeing an in-focus subject.

The rate of "looking" by the AF sensor has been estimated as in the range of 3 to 20Hz.

The "Is It Different Now" Question

Is it possible that things have fundamentally changed since RDKirk wrote about the Canon 20D? Yes, but unless he did tests like these, he would have reached the same conclusion with any contemporary body. The Squirrel Test and Intermittent Illumination Test have been verified with the 20D (see Verified Bodies).

RDKirk has been invited several times through forum posts, personal messages, and e-mail, to try the tests and comment on the results, but has not done so.

The Canon EOS 650 35mm film camera was the first Canon phase-detection AF SLR. It introduced the EOS interchangeable lens system in 1987, and has a single horizontal AF point. It behaves exactly the same as our contemporary DSLRs. See Verified Bodies.

Roger Cicala in Autofocus Reality Part 3B: Canon Cameras claims that, "The wording of this patent, back in 2003, suggests that closed-loop was not how AF worked at that time. It was largely open loop. The camera took a measurement and told the lens where it should go. Done." Roger has misunderstood the context of the Ishikawa patent, which is the control system within the lens that receives focus-change commands from the body and directly drives the focus motor to achieve those changes. Roger has confused that with the control system in the body that sends focus-change commands to the control system in the lens. It doesn't matter for focus confirmation whether the control system in the lens is open or closed-loop, since the control system in the body determines when focus should be confirmed based on what the AF sensor sees.

The "A Closed-loop Must Jitter" Misconception

Having proven to themselves that in a focus-locking mode with a half-press of the shutter button the AF system always takes more than one look, the open-loop proponents generally say, "But if it's genuinely closed-loop, how come it doesn't jitter all the time?"

A closed-loop system will not jitter (oscillate over a small amplitude around the point of focus) as long as the ball can drop into the cup, in other words, as long as the lens can position the focus within a range of acceptable focus. The Canon EF 50mm f/1.8 crude mechanism myth comes into play here as well, but the fact is, that lens can position the focus with extreme accuracy (very much tighter than the depth of field), and the source of its unreliability lies elsewhere (stay tuned for an article specifically about that). If such a system would jitter, we would see that behaviour in continuous-focus modes with good static targets in good light, but we don't.

The "Only an Open-loop Can Have Errors" Misconception

Many people get stuck on the idea that the only way such a control system could consistently produce errors is for it to be open-loop. If you assume that the AF sensor always delivers a perfectly accurate and unambiguous measure of phase difference, which results in a single point of required focus, then focus errors must be caused by the lens not going where it was told. But that would only be true if the feedback to the AF system was from the image sensor, not the separate AF sensor.

All we need to get a persistent focus error is for there to be a mismatch between a cancelled phase difference on the AF sensor and focus on the image sensor. One obvious source is a difference in the length of the light path to the image sensor and the AF sensor. That can be corrected by body calibration using a reference lens.

Another obvious possibility with Canon lenses is an imperfect Best Focus Correction Value (BFCV). (I assume the systems for other manufacturers have an equivalent feature.) Each lens communicates BFCVs to the body, which tell the AF system how far it needs to offset from a zero phase difference to account for the peculiarities of that particular lens, especially the influence of spherical aberration (see page VI-7 of Canon's EF 50mm 1.8 Service Manual), using light coming from an annulus within the exit pupil. My understanding is that the BFCV is altered in lens calibration, and something very similar is happening with AF microadjustment (see Canon EOS Autofocus – BFCV and micro-adjust (MA)).

The simple fact that lenses can and must be calibrated or adjusted tells us that persistent errors are a normal part of either type of system. If an open-loop system requires calibration to make focus on the AF sensor correspond with focus on the image sensor, so does a closed-loop system, because the accuracy achieved by both systems depends primarily on the ability of the AF sensor to measure the phase difference, and the ability of the AF system to compensate for confounding factors like spherical aberration.

The "Genuinely Closed-loop" Misconception

People who haven't studied control system theory often get confused about what "closed-loop" means. They insist that an AF system can only be "genuinely" closed-loop when the image sensor provides the feedback, as is the case with contrast-detection (CD) AF. That's simply wrong.

Closed-loop only means that the control system makes active use of information about the thing it's trying to control. There is no requirement that feedback must be a direct and perfectly accurately representation of the true state of the thing being controlled. For instance, the cruise control system in a car doesn't use measurements of the actual speed of the car, it uses the rate of wheel rotation, which is a proxy for the actual speed, and requires calibration for accuracy. Sending your mother an e-mail telling her how to put a shortcut on her Windows desktop is open-loop, whereas coaching her through it on the phone is closed-loop. In that case her description of what she's doing and seeing is a proxy for what's actually happening. In the case of PD AF, the AF sensor is a proxy for the image sensor.

The "Closed-loop = Infinite-loop" Misconception

Another source of confusion is the misconception that "closed-loop" means an infinitely repeated execution loop, like a computing "for loop" or "while loop" with no halt condition. It's true that you would use execution loops in the software which implements a closed-loop control process, but the "closed" part of closed-loop is about feedback of information, not repetition of execution of instructions. Having a halt condition in an execution loop, like "stop when the subject is in focus", doesn't mean there is no feedback loop.

The "Lens Feedback" Myth

At this point the open-loop proponents usually go, "Yeah, but the feedback is from the lens telling the body that it has completed the focus movement, not from the AF sensor. Chuck Westfall said so, right?" (Even though the Intermittent Illumination Test proves that with a half-press, focus can only be confirmed when the AF sensors sees an in-focus target.)

If focus confirmation depends on the lens telling the body that it has completed a movement, then at the instant in which you bring the subject into focus in manual focus mode, the body must be sending the lens a command to move zero, and the lens must be responding that it did. That's ridiculous – if the body knows the lens doesn't have to move, then the body doesn't have to wait for the lens to confirm that it didn't move. The rational explanation is that focus is confirmed when the AF sensor sees an in-focus subject.

The Manual Focus Test

How is focus confirmed when the AF system isn't telling the lens to change focus?

Scenario – focus is achieved by means other than the focus motor under the control of the AF system.

HypothesisPrediction
Open-loop Focus is confirmed when the lens tells the body that it has completed a focus movement.
Closed-loop Focus follows the target and is confirmed when the AF sensor sees that the target is in acceptable focus.

Method

  1. Set up a target with the camera on something like a macro slide in a focus-locking mode (One-Shot or AF-S) using a single AF point.
  2. Prefocus the lens at the minimum focus distance (MFD), AF and take the shot.
  3. Repeat step 2 nine more times and evaluate the shots to establish a typical focus under those conditions.
  4. Prefocus the lens at infinity, AF and take the shot.
  5. Repeat step 4 nine more times and evaluate the shots to establish a typical focus under those conditions.
  6. Switch to manual focus mode (MF).
  7. Set up the plane of focus just in front of the target, half-press and hold the shutter button, and carefully wind the camera towards the target until focus is confirmed*. Take the shot.
  8. Repeat step 7 nine more times and evaluate the shots to establish a typical focus under those conditions.
  9. Set up the plane of focus just behind the target, half-press and hold the shutter button, and carefully wind the camera away from the target until focus is confirmed. Take the shot.
  10. Repeat step 9 nine more times and evaluate the shots to establish a typical focus under those conditions.
  11. Compare the typical shots.

* In MF, focus confirmation is the same as in a focus-locking mode (typically, the camera flashes the active AF point, turns on the focus confirmation light in the viewfinder, and beeps when the subject is in focus), except that in MF shutter firing is not disabled to begin with, so the final state is functionally identical. Detecting focus confirmation is harder with Nikons and Sonys because they don't beep in MF mode.

Does focus confirmation depend on the lens completing a focus movement, or on what the AF sensor sees?

Bonus Questions

  1. Repeat steps 6 to 10 but turn the focus ring on the lens rather than move the camera.
  2. Take ten shots in a continuous-focus mode (AI Servo or AF-C) and compare a typical shot with those from the previous steps.

Interpretation of Results

  • No matter how the target comes into focus (either the lens changes the focus position according to a command from the AF system, or the camera-to-target distance changes in MF, or the photographer turns the focus ring), the resulting focus and focus confirmation are essentially identical if those actions are done carefully under controlled conditions.
  • The only common factor in focus confirmation is the AF sensor's view of the target.

Focus Search

When the AF sensor is unable to measure a phase difference (we are on the tee), the lens starts racking through its focus range, trying to find a position from which the AF sensor can get a measurement (trying to get on the green). Canon calls this behaviour "focus search". If you're not familiar with it, try autofocussing with the lens cap on, and you'll see focus traverse its full range.

This "detection zone" (on the green), in which focus search is not required, can include the entire focus range, MFD to infinity, with short focal lengths, and can be quite tight around the subject with longer focal lengths.

Is focus search a distinct mode of operation, or can we make sense of it as part of what the tests tell us about how focus-locking modes works?

We know that if focus is inside the detection zone when AF is triggered (we are on the green), the lens is told to head for the focus position calculated from that measurement (take a putt). As the lens gets closer to focus the continuous measurements and calculations of desired focus position get more accurate (more putts), until they converge on an acceptable focus (in the cup). Of course we don't actually putt in golf while the ball is rolling, so with continuous illumination the play becomes more like hockey.

If focus is outside the detection zone when AF is triggered (we are on the tee), all we have to do is tell the lens to head in the direction most likely to find a target, like, "if focus is near the MFD, head for infinity..." (drive off the tee). As soon as focus enters the detection zone (on the green), the process can continue exactly as it does in the first case. (This is all closed-loop of course, because during focus search we are continuously monitoring the feedback from the AF sensor to know when a measurement is available to act on.)

So I see focus search as like, "we don't have a measurement of phase difference (we can't aim to actually sink the ball with this shot), so let's just have a guess, act like we do, and see what happens (aim for the flag)." I can't see any justification for thinking of focus search as a distinct mode, only as a different initial condition to the normal process.

With a Canon body, if focus search fails to achieve a measured phase difference, the AF system will wait indefinitely and will execute another focus search if it sees something different.

Finding the Detection Zone

If you have the right gear it's easy to test whether the current focus is inside the detection zone (on the green).

With a Canon nD or nnD body you can disable focus search using the Lens drive when AF impossible custom function. Then half-press on the target – if nothing happens, you're outside the detection zone.

You can find where the detection zone starts by approaching or backing off from the target, zooming out, or (with a Full Time Manual focus lens) manually focussing towards it, and there will come a point at which the AF system takes over and achieves focus – that point is the limit of the detection zone.

Tests with a 70-300L on a 60D suggest that the detection zone may include the entire MFD-to-infinity focus range, for all target distances, for all focal lengths up to about 110mm. In other words, it appears that focus search is never required if the focal length is less than about 110mm. For a 100-400L on a 7D using spot AF, the cut-off is about 170mm. The 100L Macro's detection zone goes from 1m to infinity.

With advanced Nikon bodies, you can delay the onset of focus search after an initial response (using the Focus Tracking with Lock-On custom setting), but not disable it. But there's another way – switch to MF, half-press on the target, and if the focus indicator in the viewfinder flashes regularly, you're outside the detection zone.

The Ultimate Squirrel Test

You can eliminate focus search from the AF system's response by taking care to put your targets within the lens's detection zone (see Finding the Detection Zone). If you repeat the Squirrel Test that way (also the Intermittent Illumination Test), you should find no significant difference in the behaviour of the AF system.

Why Not Ask Canon Directly?

I did. Raymond the Online Support Supervisor replied by e-mail to a question I sent to Canon support –

"When auto-focusing the camera looks at the output of the AF sensor. Based on that output the camera knows which way to drive the lens to bring the subject into focus. The camera then sends the command to the lens to begin focusing, as the lens is focusing the AF output is monitored by the camera. Once the AF sensor detects the subject is in best focus, the camera... signals the user with the conformation beep and the shutter can be released. After AF is achieved the camera doesn't refocus unless the AF switch is released and pressed again. It's not one single measurement, as the lens is driving the camera looks at the AF output until focus is achieved."

There's no doubt Raymond is describing a closed-loop process, as is revealed by these tests.

The Phase-detection Process in Action

hpjfromdk and Rainer Hönle have captured and analysed the communication between EOS bodies and lenses, and so far their findings are consistent with the results of all of the above tests.

Using all of the information we have, we can describe how the AF control system works in a focus-locking mode, with a half-press of the shutter button, using a long lens (a short focal length lens always starts "On the Green" rather than "On the Tee").

On the Tee (click on the image to see it full-size)

The blue lines show how the phase difference increases either side of the current focus position.

When AF is triggered, the camera asks the lens for parameters such as maximum and minimum aperture, current focal length, and current focus position. The calculation of required focus position includes the phase difference seen by the AF sensor, the appropriate Best Focus Correction Value (BFCV) provided by the lens, any other adjustments set in the body, and the Focus Sensitivity Coefficient which tells the body how to convert a corrected and adjusted phase difference into a meaningful command to the lens (see Canon EOS Autofocus – BFCV and micro-adjust (MA)).

It seems there is a BFCV for each of a number of AF point groups, for each level of autofocus precision, for each discrete combination of focal length and focus position. Most Canon bodies have two levels of precision, f/2.8 and f/5.6, and some high end bodies have f/8 arrays as well. A typical zoom might encode 16 discrete focal length values over it's range, and 8 discrete focus position values. So the total number of BFCVs for a typical zoom is probably in the order of several thousand.

In the above "on the tee" case the phase difference at the subject is too wide for the AF system to determine a required focus position (we're on the tee), so when AF is triggered the AF system has a guess and tells the lens how fast to go and to head for infinity (we drive off the tee, aiming for the flag).

The AF system will "look" for the subject any time the focus motor speed is constant (which includes speed = zero). So for focus search, the AF system will start looking for a measurement of phase difference once the lens tells the body that the focus motor has achieved the required speed.

On the Green (click on the image to see it full-size)

When focus enters the detection zone around the subject a required focus position can be calculated (we're on the green), and a movement to a specific destination can be commanded to the lens (we take a putt).

There's no guarantee that the calculation will give the exact required focus position. Some lenses do need just one putt most of the time on a static subject, while others need multiple putts if they are more than a little out of focus. That's not a problem for a well designed closed-loop process, as long as the calculation can converge on the real required focus position when the focus error is small.

The AF system continues to request BFCVs and other relevant values, and to calculate and update the required focus position. With continuous illumination the AF system doesn't have to wait for a focus motion to complete (we don't have to wait for the ball to come to rest before hitting it again), so it's like dribbling in hockey. That continues until the focus motor needs to decelerate to achieve the required focus position.

Once the focus motor is at rest, if the AF sensor doesn't see an in-focus subject, fresh speed and destination commands are issued (we keep putting, as shown by bonus question 2 of The Intermittent Illumination Test). That explains the twitch at the end of focus motion, which is erroneously thought to be the lens correcting for its inability to stop on the required focus position, but the captured communication between the body and the lens shows that it is the body correcting for its inability to accurately specify the required focus position from far away.

 In the Cup (click on the image to see it full-size)

When the AF sensor sees that the current focus position is within the acceptance zone around the subject (we're in the cup), the AF system will confirm focus in a focus-locking mode (no more putts on this hole).

The acceptance zone just needs to be larger than the smallest step that the focus mechanism of the lens can make. I have tested the Canon EF 50/1.8 II and found that the acceptance zone for that lens is about 3mm deep for a target near the MFD, whereas the smallest step that lens can take is something less than 0.1mm – that's 5% of the thinnest depth of field, 3% of the acceptance zone, and 0.02% of the MFD.

My testing with Canon bodies shows that PD AF always ends up on a boundary of the acceptance zone, usually on the far side with respect to the direction of approach. In other words, focus confirmation seems to be triggered by leaving the acceptance zone rather than entering it. With MF, focus confirmation will begin at the nearest of the same two boundary points, and continue throughout the acceptance zone. A conventional depth of field (DOF) calculation for the above scenario with the EF 50/1.8 gives about 4mm for f/2.8, so it's easy to see how a focus at one end of the acceptance zone can be perfect while one at the other end (3mm away) it can be outside the f/2.8 DOF.

With a full-press of the shutter button on a Canon body, there is a zone slightly wider than the acceptance zone from which a final movement does not require a subsequent look to check focus accuracy.

In a continuous-focus mode, the AF system continues monitoring the phase difference and will take action if the focus falls outside the acceptance zone. The AF system can predict the focus position required when the shutter will open by calculating a trajectory for a moving subject from a series of recent phase differences.

The Canon 5D Mark II and 50D White Paper tells us that, "Focusing is started using the f/5.6, horizontal line-sensitive central sensors... When focusing is nearly achieved (and when using an f/2.8 or brighter lens) focusing is handed off to the f/2.8 sensitive vertical line sensors." (p. 32.) So we know that the process can involve measurements from multiple AF sensors at the same AF point.

Pseudocode Model

All of the above behaviour with a half-press of the shutter button is described by the following pseudocode model, in terms of the phase difference as seen by the AF sensor being "acceptable" or not.

(Sorry there's no indentation, the editor keeps discarding it.)

function phase detect ()

if AF and focus-locking mode

disable shutter firing

else

enable shutter firing

end-if

while true

if current phase difference is acceptable

if MF or (AF and focus-locking mode) // confirm focus

beep
flash AF point
focus confirmation light = on

end-if

if AF and focus-locking mode // halt

enable shutter firing
return

end-if

else // current phase difference is not acceptable

focus confirmation light = off

if AF

while current phase difference is indeterminate or out of range

rack lens to see if there is a phase difference somewhere
focus confirmation light = blink
wait for current view to change significantly

end-while

if focus-locking mode

calculate lens command from current phase difference

else // do predictive focus

calculate lens command from current & previous phase differences

end-if

update lens command

end-if // AF

end-if // current phase difference is acceptable

end-while // true

end-function // phase detect

Focus Accuracy

From these tests and many others with Canon gear, I have made the following tentative conclusions about the accuracy of the focus achieved by a PD AF system –

  1. Accuracy of focus depends on the AF sensor's ability to determine when the phase difference is within acceptable limits.
  2. Accuracy depends on the state of calibration of the body and the individual lens (via the Best Focus Correction Value, AF microadjustment, and the Focus Sensitivity Coefficient).
  3. Accuracy depends on any direct effects of optical imperfections which make it harder for the AF sensor to determine when the phase difference is cancelled, for instance, spherical aberration may cause asymmetrical bokeh, leading to different focus depending on whether the plane of focus started in front of or behind the subject.
  4. The effects of spherical aberration can be reduced by masking the outer annulus of the lens (where the aberrant light comes from), or by forcing the camera to use a low precision AF sensor (which is not illuminated through the outer annulus of the lens).

In other words, the presence of focus errors, which is used to support the idea of an open-loop model, is due to the imperfection of the interaction between the lens optics and the AF sensor.

See also Variation Facts and Fallacies.

Okay, So it's Closed-loop – Why Should I Care?

Knowing that it's a closed-loop system shouldn't make any difference to how you operate your camera or the results you get – good techniques are effective because they work with the AF systems we have, not because they should work with AF systems we don't have. But knowing might help you to recognise bad techniques.

When you're trying to solve focus problems, the wrong understanding will lead you to the wrong answers. For instance, people are advised to AF repeatedly, which emulates an iterative closed-loop system. This makes perfect sense assuming an open-loop model. But if you test that carefully under controlled conditions (e.g. Autofocus Reality Part 2: One vs. Two, Old vs. New), you will find that it makes no difference, because once focus is confirmed (in the cup) it will continue to be confirmed in exactly the same place, unless something happens to cause the AF sensor to see the subject as out of focus (back on the green).

Verified Bodies

These tests have been executed using the following bodies:

 SquirrelUltimate
Squirrel
Intermittent
Illumination
Open-loop
Final Step(1)
Manual
Focus
Canon EOS 650(2) Verified Verified Verified N/A Verified
Canon EOS 1Ds III Verified        
Canon EOS 10D Verified     N/A  
Canon EOS 20D Verified   Verified N/A  
Canon EOS 40D Verified Verified Verified Verified Verified
Canon EOS 400D Verified   Verified N/A  
Canon EOS 450D Verified Verified Verified N/A Verified
Canon EOS 50D Verified   Verified   Verified
Canon EOS 60D Verified Verified Verified Verified Verified
Canon EOS 7D Verified        
Nikon D200 Verified     N/A  
Nikon D7000 Verified   Verified(3) N/A  
Nikon D90 Verified   Verified N/A Verified
Olympus E-5 Verified        
Olympus OM-D E-M1 Verified        
Pentax K-5 Verified   Verified    
Sigma SD10 Verified        
Sony Alpha A230 Verified   Verified N/A Verified
  1. The Open-loop Final Step Test requires Focus Search to be disabled. The table shows where that is not applicable (N/A).
  2. The Canon EOS 650 35mm film camera was the first Canon SLR with phase-detection AF. By behaving in the applicable tests like a contemporary DSLR it proves that EOS has always been a closed-loop control system.
  3. Apparent open-loop behaviour has been reported with a Nikon D7000 when using AF modes outside the scope of the test, such as 3D-tracking.

Conclusion

These tests prove –

  • If the shutter button is half-pressed in a focus-locking mode, focus is confirmed if and only if the PD AF sensor sees something in acceptable focus.
  • Feedback from the PD AF sensor is effectively continuous in good light.
  • Focus-locking PD AF modes are closed-loop control processes, just like continuous-focus modes.
  • PD AF in Canon EOS cameras has always been a closed-loop process.
  • For a Canon body, if the shutter button is full-pressed in a focus-locking mode with the subject just out of focus, the shutter can fire after a single focus movement without a subsequent check of focus accuracy via the PD AF sensor.

The tests also show that with static subjects, the only difference in behaviour between a focus-locking mode and a continuous-focus mode is that a focus-locking mode halts the process when focus is acceptable. (Continuous-focus modes have a predictive element with moving subjects, but that is not a factor in these tests.) So we can infer and assume that the focus-locking mode and the continuous-focus mode within a given camera are a single closed-loop process, with the addition of a halt condition in the focus-locking mode (and a predictive element in the continuous-focus mode).

It's reasonable to assume that focus search is the AF system operating normally using a guess of where focus might be, in the absence of a measurement and calculation of where it should be.

The open-loop model of PD AF is proven to be a myth, based on incorrect or misleading assertions and its apparent fit with cursory observation of the behaviour of the PD AF system, and the closed-loop model is confirmed by Canon support and Scientific American. In the voice of Mythbuster, Adam Savage, it's "Busted!"

If you don't understand any of the above, or disagree with it, play around with the tests and verify their results before you leave a comment. You really have to experience it for yourself to break through the myth. And please let me know if you run the tests with a body other than those already listed.

Thanks to Doug Kerr for all his hard work, patience, critical thinking, and the golf analogy, to hpjfromdk for decoding body-lens communication, to imqqmi for the idea of the diagram used in The Phase-detection Process in Action, to Catherine for the loan of a D90 and Colleen for the A230, and to all the commentators who have done independent verifications and helped me make this article better.

Further Reading

Doug Kerr's Principle of the Split Image Focusing Aid

Canon's EF 50mm 1.8 Service Manual (see page VI-7 about calibration via the BFCV)

Doug Kerr's Canon EOS Autofocus – BFCV and micro-adjust (MA)

Stanford CS 178 - Digital Photography Autofocus: Phase Detection

Control Theory

The views and opinions expressed in this article are those of the author and do not necessarily reflect the views and opinions held by dpreview.com or any affiliated companies.

Comments

Total comments: 41
kmanoz
By kmanoz (May 24, 2012)

Wilba, love your work. New to photography, but an Engineer of 17 years and I can't fault your analysis.

Having read countless statements to the effect that an AF micro adjustment completely solved a given users front-focus condition, for example, I was at a loss to understand how a system that simply has to be closed-loop could possible return a consistently incorrect outcome. Of course, knowing that a lens has a parameter called BFCV, it's all too obvious.

One of the things that makes finding information on this subject so very difficult is that the vast majority of users of this technology don't seem have any scientific/engineering skills. And so, like religion, the real truths of the matter can be completely obscured by the word of the masses. I still have faith that the Internet can survive as a powerful tool of knowledge and wisdom when such scripture as this exists.

Now to spread the word. :)

2 upvotes
WilbaW
By WilbaW (May 28, 2012)

I agree with everything you said! :-)

Thanks a bunch, Wilba (B. Eng. Mech.)

PS: If you're new to photography you might get something out of my other articles.

1 upvote
TomFL
By TomFL (May 17, 2012)

Good article. Well written.

As to the question of why an engineer may choose a known inferior method, the answer in today's world is unfortunately PATENTS.

The US patent system is now polluted with an enormous number of bogus patents, especially software patents. Seemingly obvious methods of feedback related control get patented and a competitor can lock you out for 20 years.

And the irony is the patent system was created to foster innovation. If you have been in engineering more than 10 years, you will have probably been bitten by this already.

1 upvote
WilbaW
By WilbaW (May 28, 2012)

Thanks Tom. Good point with patents, though in this case we know that cameras from three different manufacturers operate according to the same principle, so I guess it's not restricted that way.

0 upvotes
WilbaW
By WilbaW (Nov 2, 2012)

Now _four_ different manufacturers.

0 upvotes
WilbaW
By WilbaW (8 months ago)

Apparently Honeywell patented PD AF in the 1970s. They sued Minolta for patent violation and won, and the other manufacturers signed up to pay royalties for the system rather than face the same battle. That makes sense of the observation that the systems from the four manufacturers tested so far behave fundamentally the same.

0 upvotes
WilbaW
By WilbaW (4 months ago)

Make that five - Canon, Nikon, Olympus, Pentax, and Sony (Minolta). Is that everyone?

0 upvotes
WilbaW
By WilbaW (4 months ago)

Sigma too.

0 upvotes
Amateur Hour
By Amateur Hour (Feb 3, 2012)

Great article. Having worked as a software engineer for an electronics manufacturer, it's very unlikely that an engineer would choose an open-loop system, when you can get better performance with less rigorous requirements in a closed loop. Open loops require precise estimation, and would also require that the electro-mechanical focusing mechanism also be calibrated in terms of travel. With a closed system, you can use a PID calculation to get it very close on the first reading, and nail it on the second or third try - not to mention that the focus is verified as well.

Thar fact that system can fail is not an indictment - one the system determines it is within focus tolerance, it will stop in order to prevent oscillation. loops designs *may* continue to read the sensor - but they may only make adjustments once the reading goes outside focus tolerance. Careful combinations of this kind of logic make for features such as one-shot and servo-ai focusing modes.

Comment edited 6 minutes after posting
0 upvotes
WilbaW
By WilbaW (Feb 4, 2012)

Yeah, it's obvious when you have some understanding of control system engineering. Thanks very much for commenting.

1 upvote
speedmasterMK4
By speedmasterMK4 (Jan 14, 2012)

I think I started a thread about this in 2005 trying to bust the open loop myth but the discussion got ugly. I am glad someone has taken the time to write the nice article and clear things up.

Thanks

0 upvotes
WilbaW
By WilbaW (Jan 19, 2012)

Thanks. Yeah, it always gets ugly when you challenge people's beliefs. :- )

0 upvotes
SteveL54
By SteveL54 (Oct 29, 2011)

Thanks for the fine article

I just completed intermittent illumination test on Canon 20D with EF 50mm f/1.4 lens. Used a Nikon SB-800 off camera as the light source. The target has very bold black & white lines of irregular width running horizontally. The center AF point was used.

I set up as you specified except the camera was almost 6 feet away from the target (not very near MFD). I could not disable focus search.

I repeated the steps several times. In all cases, when I had the room very dark, two flashes were needed for focus confirmation. ( At higher room lighting level -- but still low -- I had one test cycle that took three flashes. Another cycle took just one, but it's photo was OOF.)

On a few of the cycles I looked at the lens distance scale after the first flash and it had always reached the correct focus distance. On second flash focus confirmation was made and if I had the button fully depressed it took the photo (exposure was several seconds at f/2 and ISO 800)

Steve

0 upvotes
WilbaW
By WilbaW (Oct 29, 2011)

Thanks very much Steve. Everything you observed fits with what I would expect.

Did you try the Squirrel Test?

0 upvotes
SteveL54
By SteveL54 (Oct 29, 2011)

With the 20D I wasn't sure what to expect. I don't use it as much as my Nikon gear.

From tests carried out a few years ago on my Nikon D70 and D300 I was sure they are closed loop under low light conditions or if the target is particulaly challenging. The multiple movements under these conditions are fairly obvious.

I haven't tried the Squirrel Test. (It was too dark to take advantage of the long focus travel distances outdoors. Besides all the neighborhood squirrels were asleep.) The Squirrel Test can test the system under normally bright light conditions when the system *might* normally operate in a mode without the terminal focus confirmation measurement.

I don't know if I can get to it this weekend, so if anone else with a 20D and the time is curious I'd be interested in their results as well.

Steve

0 upvotes
SteveL54
By SteveL54 (Oct 29, 2011)

I had time early this afternoon and was curious about the Squirrel Test. The AF behaved as you would have expected.

As soon as I initiated AF on the near subject I could see the AF moving into focus on the near subject. But right after quickly panning to the background, focusing direction reversed and stopped on the background with confirmation.

I tried it several times the other way with consistent behavior. Focus would start to move towards the background but reverse towards the forground target once it was swung into the center AF point.

Both near and far targets where directly illuminated by the sun. Near target about two yards away while far target was about 20 yards. Near target a tree trunk with good texture and the far target a mix of brush and small tree trunks. The 20D AF system reliably and smoothly aquired focus lock if I dwelled on the targets so they are both good targets.

I used an EF 50mm f/1.4 lens and a 100-400 L IS. (IS off)

Steve

0 upvotes
WilbaW
By WilbaW (Oct 30, 2011)

Thanks very much Steve. Yes, there's no doubt you're describing the operation of a closed-loop control system.

I'm going to add a section entitled The Ultimate Squirrel Test to more explicitly address a valid criticism of the basic ST.

1 upvote
Fred Mueller
By Fred Mueller (Oct 26, 2011)

Any idea why " Nikons need a lot longer" in the intermittent test?

And if so, would that be related to the general observation that Canon AF is generally faster than Nikon, but Nikon seems to be more accurate? (putting on flame proof jacket)

I currently shoot a D700 (very good AF) , have extensive experience with a D300 (also good, but a little slower) and D40 (slow but very accurate, especially in low light). Had a D7000 for a few months, it had AF "problems".......

0 upvotes
WilbaW
By WilbaW (Oct 27, 2011)

Here's what I observed - it only takes a low-power flash to make a Canon system respond, whereas no matter how much flash power I used with the D90, it would not respond. I had to open a door to a brightly-lit adjacent room, or switch on a light, to get the D90 to start AFing. So it's not the intensity that matters, it's the duration of illumination.

Apart from that, the two systems seemed to perform pretty much the same (with lenses with their own motor), so I assume the lack of response from the D90 is a policy rather than reflecting some fundamental difference in hardware or operating principles.

I can't comment on the relative merits of the two systems :-), but it did occur to me that the dedicated AF assist light on Nikon bodies would make good sense - if the policy is to not respond to very brief illumination and use AF assist instead.

Would you be willing to try at least the Squirrel Test and the Intermittent Illumination Test with the D700 and let me know how you go?

Comment edited 1 minute after posting
1 upvote
Fred Mueller
By Fred Mueller (Oct 27, 2011)

Really don't have the time in the next few weeks - up-to-eyeballs in work load.

Do really appreciate your article above - thanks for the insight....

0 upvotes
dopravopat
By dopravopat (Oct 20, 2011)

Perfect, thanks. Your article explained it very well and I would recommend it to every photographer. Understanding the process of autofocus enables to help the camera focus in bad light, to get more consistent results and also to use properly focus confirm in MF mode.

Comment edited 9 minutes after posting
1 upvote
bronxbombers
By bronxbombers (Oct 18, 2011)

one thing i wonder is why it happens often enough that it will badly miss focus, way beyond any focus calibration mistake level and yet report prefect focus and sometimes not a single spot on the entire image is in focus and even with an easy subject and condition

0 upvotes
WilbaW
By WilbaW (Oct 20, 2011)

Yeah, I have seen that in unusual situations. In normal operation, focus is confirmed when the AF sensor sees an in-focus subject, so I guess it's possible for the AF sensor to be fooled - to see a pattern that looks like a cancelled phase difference with an out of focus subject.

0 upvotes
Dale Buhanan
By Dale Buhanan (Oct 18, 2011)

I became convinced that it was more than a "one look, one movement, open loop" design during the discussion we had some time ago on DPR. I do agree with your observations here, since I duplicated them also. It is clear to me that there is closed loop behavior, although I don't see it as a completely closed loop design in the classical sense.

It is more like a hybrid closed loop approach, or what I would calll a 'piece-wise linear' closed loop, so that oscillation about the focus point cannot occur. ie., the loop doesn't remain closed during the focus process, it is briefly interupted (opened) for evalutation at the end of each movement. That prevents oscillation, and is the best compromise between speed and accuracy.

Excellent work, WilbaW. Good rational thinking, which has furthered all of our understandings on the AF mystery. I am better off with better understanding as a result of your work and my discussions with you. Thanks! -Dale

Comment edited 4 minutes after posting
0 upvotes
WilbaW
By WilbaW (Oct 19, 2011)

Thanks Dale. Which body did you test?

Piece-wise linear behaviour is apparent with intermittent or very low illumination, but the Squirrel Test shows that a focus command doesn't have to complete before the lens can head for a new target. In other words, when there is sufficient illumination, the rate of phase measurements (estimated to be between 3 and 20Hz - http://forums.dpreview.com/forums/read.asp?forum=1032&message=31125414) and the system's response to them is faster than the rate of completion of focus movements.

0 upvotes
Loren Charif
By Loren Charif (Oct 18, 2011)

Although I haven't read ALL the member-written articles on DPR, this is by far the best I have read. Well written, clearly explained, logical, and relatively easy to understand, even for those technically-challenged brains like mine.

Thanks!

Loren

0 upvotes
WilbaW
By WilbaW (Oct 20, 2011)

Thanks very much Loren. I wrote it for you. :-)

0 upvotes
Kasanicky
By Kasanicky (Oct 18, 2011)

Yes I agree with you, I have already noticed it, and this is reason I am now using back button focus. If I am in focus and release back button, during taking picture it is not refocusing any more.
Thank you for interesting explanation, very nice post.
Kind Regards
Karol

0 upvotes
aVolanche
By aVolanche (Oct 15, 2011)

Interesting and informative read, Wilba. I'd have to agree with you, based on a gut-level feeling. I need to repeat your tests, but am sure I'll get identical results.

Has this changed your approach to using your camera in any way?

Fred

0 upvotes
WilbaW
By WilbaW (Oct 16, 2011)

Thanks Fred. Please let me know how you go with the tests.

I can't think of anything I really _do_ differently, since both models lead to the same observable behaviour in most cases, but I think it helps to understand how it works when making choices between various focus options and techniques.

I often see the myth leading to the wrong advice when people try to solve focus problems, so the main purpose of this article is to have a resource to refer to.

The bumper sticker version is, "focus is confirmed when the AF sensor see an in-focus subject". Let me know if that changes anything for you. :-)

0 upvotes
aVolanche
By aVolanche (Oct 16, 2011)

No, it changes nothing for me either. Your article, along with Rudy Winston's video at B&H, gives a great look at Canon AF (all the "hows" and "whys"). Nice work!

Regards, Fred

0 upvotes
WilbaW
By WilbaW (Oct 15, 2011)

Thanks janzu. For a long time I've wanted to put all this information in one place, and the Articles seemed like a good way to do it.

0 upvotes
janzu77
By janzu77 (Oct 14, 2011)

Thanks Wilba for thorough explanation! Lot's of interesting things there to read :)

0 upvotes
tomlang
By tomlang (Oct 20, 2011)

Great article, but it leaves me with one question: What about all the reports of people complaining about front focus and sending the camera and lens back to Canon for re-calibration? Just what is re-calibrated?

0 upvotes
WilbaW
By WilbaW (Oct 20, 2011)

Thanks Tom. Canon don't explain what happens in calibration, so I don't know much more than what I said about it in the Counter Arguments section of the article.

Comment edited 9 minutes after posting
0 upvotes
mosswings
By mosswings (Oct 24, 2011)

The light path to the PDAF sensor is different than that to the main sensor, so the focusing system has to be adjusted to account for this difference. Sensor calibration consists of electronically and mechanically matching the path length and sensor orthonormality. This is perhaps the only true "open-loopedness" of the PDAF system - it never actually uses the main sensor (it physically can't in moving-mirror SLRs).

0 upvotes
takya
By takya (Nov 26, 2011)

Great, deep article, but one more point that might be added, in addition to mosswings remarks: One aim of continuous (PD)AF is to sense the motion of the target and predict its future location in order to cancel out the delay of the AF system so that the lenses are always already at the correct position whenever the shutter is pressed (minus shutter delay), so, for any moving target it's actually the continuous AF that is inherently open-loop, and only a one-shot AF can be a closed-loop. ;-)

Wished that camera manufacturers would build a real closed-loop one-shot PDAF using some self-calibration via CDAF though. :-)

Comment edited 3 times, last edit 3 minutes after posting
1 upvote
WilbaW
By WilbaW (Nov 26, 2011)

Thanks takya. Continuous AF (predictive or otherwise) is by definition impossible for a "one look, one movement" system to do, so I assume you're confused about what "open-loop" means (the article covers all the misconceptions I can think of).

mosswings' comment is covered in the section on The "Genuinely Closed-loop" Misconception.

I agree, there's no reason I can think of why Canon can't implement a self-calibration system that tunes PD AF using CD AF for reference. I want a button to trigger it. I don't mind if it takes five seconds. :- )

Comment edited 53 seconds after posting
0 upvotes
cgarrard
By cgarrard (Jun 1, 2012)

Totally agree. Self calibration of lenses should have been introduced when CDAF LV was introduced into DSLRS. It does mean however that shops/manufacturers lose out on the revenue generated by such service rendered. Probably why we don't have it yet.

In the world of DSLRS and the point of diminishing returns I'm starting to see- its features just like this that manufactures need to start thinking of letting the end user have control of.

Excellent article too Wilba, thoroughly enjoyed every word of it. Just waiting for RDKirk to reply now. (Grin).

Carl

0 upvotes
WilbaW
By WilbaW (Sep 14, 2012)

Hello Carl, thanks. Somehow your reply got hidden so I saw it for the first time today.

Sadly, RDKirk is still not responding. He was recently invited to comment (in http://www.lensrentals.com/blog/2012/07/autofocus-reality-part-ii-1-vs-2-and-old-vs-new), but nothing came of it.

I have no problem with what he said back then (it made sense and fits with what Canon have always said). The problem is that it is taken as gospel, and the true believers simply refuse to gather and examine the evidence that would prove their beliefs.

OTOH, my guess is that Chuck Westfall knows the truth but is not allowed to discuss it.

Comment edited 29 seconds after posting
0 upvotes
cgarrard
By cgarrard (Sep 14, 2012)

Yep I hear ya. It's up to the individual to find truth, when you entrust others for the truth you give up a very valuable right.

Knowledge.

Good article, good detective work (thanks for PM'ing me).

Carl

0 upvotes
Total comments: 41