Originally posted by Christine Tham Agree.

The eyes are typically low contrast with no well defined sharp edges, so the AF sensor will probably focus on something else - I only mentioned the spoon because in the photo that object is brighter than the face and has sharp edges so would register very strongly on the phase detect.

I don't know if anybody is interested in this.

Normally, what a phase AF does, is this:

it takes the two 1-D images from the two sides (left/right or top/down) of an AF sensor's line sensor. It ends up with two curves ci(x) from each sensor. The curves would have large values for bright regions and low values for dark regions (normally, ci is defined such that Integral_x ci(x) dx = 0 but it doesn't matter here).

Now, it computes the so-called cross correlation function between both curves, where one curve is shifted against the other curve by a so-called phase p: Integral_x c1(x)*c2(x-p) dx

In a quick optimization, it detemines p such that this is maximized. (Actually, this isn't fundamentally different from what a contrast AF system does. Which is why I sometimes don't differentiate between phase and contrast AF, much to people's surprise normally

)

This value of p is then used to determine the necessary lens shift to aquire focus.

The important point is this:

Whatever p maximizes Integral_x c1(x)*c2(x-p) dx is taken.

So, you cannot say the AF sensor focussed on feature x. It's not as simple as this. To further complicate things, a vendor can experiment with functions, e.g., replace them by ci^2 etc. Unmatched contrast edges make a correlation function small which is why they matter. But all contrast variations within an AF sensor's area contribute to the phase decision to a smaller or larger amount. It's not feature x like pupil or spoon.

Sidenote:

In the presence of noise, the computed value of p starts to be noisy too up to the point where it becomes a random value. An AF module has to decide how much light it receives (in order to take longer captures with its AF sensor) and if it can trust the value of p. Both are sources for mistakes being made by the firmware of the AF module if it is autonomous.