How Canon PD AF Works

Crikey, I'm excited! I just found the best way yet to look inside the
operation of Canon's phase detect auto-focus. This emphatically drives a
stake through the heart of RD Kirk's open-loop theory.

Here's the test. I call it the AF-assist Test. I used a 450D, EF-S 55-250,
and a battery-powered torch (flashlight). Set up a target at a focus-able
distance in "darkness" (not enough light for the system to achieve focus
without assistance). Use One Shot PD AF, centre AF point, and manual
exposure mode to the stop on-board flash popping up.

1. Focus the lens at a distance other than the target distance.
2. Trigger AF on the target, and hold the button down throughout the rest of
the test.
3. The system will fail to focus (no beep, focus confirmation light
flashes).
4. Flash the torch on the target (the briefest flash you can make).
5. The lens will move to the in-focus position but the camera will not
confirm focus (no beep, focus confirmation light still flashes).
6. Flash the torch again.
7. The camera will confirm focus (beep, focus confirmation light stays on,
AF point flashes red).

This test proves that focus is confirmed when the AF sensor sees that the
phase difference is cancelled, not when the camera believes that the lens
has moved to where it was told to go, as in the open-loop myth.

Under normal conditions of adequate illumination, the active feedback via
the AF sensor is available continuously, and the process occurs rapidly
enough that it generally appears seamless, just like the
one-measurement-one-movement myth.

You can do the Squirrel Test this way too -

1. Focus the lens at a distance other than the target distance.
2. Trigger AF on the target in darkness, and hold the button down throughout
the rest of the test.
3. The system will fail to focus (no beep, focus confirmation light
flashes).
4. Flash the torch on the target (the briefest flash you can make).
5. The lens will move to the in-focus position but the camera will not
confirm focus (no beep, focus confirmation light still flashes).
6. Aim the camera at a target at a different distance.
7. Go to step 4, and repeat until satisfied.
8. Skip steps 6 and 7, and flash the torch again.
9. The camera will confirm focus.

Once you accept that focus is confirmed when the phase difference is
cancelled, everything gets very simple (William of Ockham is cheering with
me). One Shot, the Squirrel Test, AF-assist, trap focus, the Beep Test (like
trap focus but with the lens in MF)... are all just the operation of a
single closed-loop process under different conditions. And AI Servo is the
same process too, except that focus confirmation is turned off so there is
no halt condition.

Inevitably, someone will respond that a closed-loop system must always
provide an optimal focus, like CD AF does, or will hunt endlessly using
phase detect, and since these systems do mis-focus with certain lenses,
without hunting, this theory must be wrong. The evidence and possible
reasons for focus errors have been discussed at excruciating length in other
threads and forums, and I'm not interested in getting into that again. So
just do the test and see what you get out of it. Feel free to disagree with
my interpretation of these results, but do that through a theory that
comprehensively explains the results you get from doing it yourself.

Enjoy! :- )

Wilba wrote:
4. Flash the torch on the target (the briefest flash you can make).

Try an external flash gun and trigger it --- that'll be even
briefer.

-Wolfgang

Wolfgang Weisselberg wrote:
Wilba wrote:

4. Flash the torch on the target (the briefest flash you can make).

Try an external flash gun and trigger it --- that'll be even
briefer.

True but unnecessary. As long as you don't shine a light throughout the
motion of the focus ring, you can break down the AF process into its
essential step.