Lenses....some Focus faster than others...

[Daniel Browning]

asmodai, I think you are mistaken; wickerprints did an excellent job explaining the situation, but I'll see if my two cents can help.

A larger lens collects more light, so more contrast information has to be potentially available. [...] optically, the wider the lens the better the AF could perform.

That's not how PDAF works. As wickerprints explained, the autofocus sensors can only see a certain part of the lens. No matter how much faster than f/2.8 the lens is, the contrast information available to the autofocus stays the exact same.


The only possible way for that to change is if a manufacturer somehow came out with the first ever f/1.4 or f/2 autofocus sensor. If that ever happened, it would be the only circumstance in which having lenses faster than f/2.8 makes any difference whatsoever.

[asmodai]

That's not how PDAF works. As wickerprints explained, the autofocus
sensors can only see a certain part of the lens. No matter how much
faster than f/2.8 the lens is, the contrast information available to
the autofocus stays the exact same.


The only possible way for that to change is if a manufacturer
somehow came out with the first ever f/1.4 or f/2 autofocus sensor. If
that ever happened, it would be the only circumstance in which having
lenses faster than f/2.8 makes any difference whatsoever.

Ahh! I think I understand the impasse here.





So, yes. it is true. Phase detection AF sensors become available in discrete intervals based on the lens they were designed for. No wider lens makes an already accessible af sensor more accessible. No argument there.





Moving beyond that, on to the AF sensor itself, onto the photosensitive surface area. They work by discerning light patterns (contrast patterns), and bringing them into alignment with one another by moving groups in the lens. Now, any given AF point represents some area in the circle of light thrown into the camera from the lens, some fixed point. The light cast on that static photosensitive area will be MORE INTENSE given identical scenes, identical focal lengths, using the same AF points, but with a wider lens capable of collecting more light. Static area, static position, but more photons hitting that surface area. More information HAS to be available. It's unavoidable, isn't it?


And in live view / contrast detect mode, forget about it. But I can imagine we'd disagree on that =]





I am open to being proven wrong if I am, so please, if I'm making some error here, let me know

[asmodai]

http://www.cambridgeincolour.com/tutorials/camera-autofocus.htm





This is a really, really well written description of phase detection AF, and it's dependence on contrast. And as I mentioned, as as they touch on, contrast is dependent on sufficient light being gathered by the lens.

[wickerprints]

You are conflating AF acquisition with AF precision, which are two distinct concepts. It is one thing to ask the AF system to find focus, and another to ask the AF system how precisely it has found focus. The former requires a certain level of subject contrast across the AF region (and along the direction the sensor is sensitive), but the latter requires a certain baseline separation and therefore requires a lens with an aperture of a certain minimum diameter.


As I pointed out regarding the EF 85/1.2L II, the AF sensor is not slowed down because it has to find a more precise focus--the f/2.8 sensor simply isn't capable of discerning the f/1.2 image. And by this, I mean the change in AF target sharpness at f/1.2 when the focusing group is moved within +/- 1/3 of the f/2.8 depth of focus. That's what I am trying to explain. The slowness of AF due to insufficient light (and therefore insufficient subject contrast) is an entirely different aspect to AF performance.


If you think a bit about AF precision and how it pertains to the rated sensitivity of the AF sensor points, this makes perfect sense. If an f/2.8 sensor were able to discriminate changes in sharpness at an f/1.2 "level," then the same could be said for an f/5.6 sensor. There is a reason why the system is engineered with both sensitivities. The less sensitive points are needed because the more sensitive points become useless for slow lenses, and the more sensitive points are needed because the less sensitive points are, well, less sensitive! What goes for f/2.8 vs. f/5.6 also applies to f/1.2 vs. f/2.8.


Coming back to required contrast then, this is something that is correlated to but not strictly dependent on light level. Yes, you need a certain minimum EV for the AF to operate. But if it can operate, it doesn't get more precise--nor does it get faster--the more contrast you have.





As for Live View / image sensor-based contrast AF, this is a completely different creature from the phase-detect AF. The LV AF works by direct software processing of the pixel readout in the chosen area of interest, and is therefore working at whatever maximum aperture the lens has. In other words, if you enabled LV AF with the 85/1.2L II, you could in theory obtain precise focus of the f/1.2 image--in practice, the focus precision is a function of the camera's contrast-detection software algorithm.

[Daniel Browning]

The light cast on that static photosensitive area will be MORE INTENSE given identical scenes, identical focal lengths, using the same AF points, but with a wider lens capable of collecting more light.

Not so. Here's how it works. Imagine you're inside the camera, standing on the image sensor. When you look out, you can see the entire lens aperture. At f/16, the light hitting you is not that intense. But as the f-number gets faster, you can still see the entire aperture, and the light gets brighter until you burn like an ant under a magnifying lens.


Now imagine you're standing on the autofocus sensor. There is a giant tube restricting you from seeing any part of the lens except the f/2.8 ring. At f/16 you see nothing. At f/2.8 you finally see something, but it's not that bright because it's only one part of the lens. Your not even burning at all. At f/2, it still looks the exact same, because the wider part of the aperture is completely blocked from your view. You don't notice at all when the f-number is changed to f/1.4, either.


The autofocus sensor gets light from only one part of the lens, so the intensity does not increase with faster f-numbers. Intensity only increases for the image sensor because the image sensors sees light coming from all angles of the aperture. But the autofocus sensor does not see all angles: it restricts itself to just one part of the aperture (as a necessary part of its design).

[MikeWhy]

Now imagine you're standing on the autofocus sensor. There is a giant tube restricting you from seeing any part of the lens except the f/2.8 ring.
<div style="clear: both;"]</div>

Now, how does that work? I'm a very small ring of photosensitive material with my view of the world occluded by an umbrella that covers the center 2*atan(1/2*1/2.8) angle of view. (Based on what depth?) Then what happens?


As the lens approaches focus, the light from the subject is scattered less and less. Its CoC on the sensor approaches zero. At some point, an edge detail will read more strongly on my neighbor than on me. So, by peaking the difference between my neighbors and myself, the AF system determines it found focus.


That's it? No weird integration or anything? Why occlude the center at all?

[Daniel Browning]

Why occlude the center at all?

The purpose is to get two or more different views of the same subject. My understanding is that it's more than just a central occlusion: each of the paired AF sensors gets light from an aerial image of a virtual subaperture at opposite sides of the f/2.8 annular ring. If it were designed to see more of the center, the accuracy would decrease in proportion with the decrease in baseline. If it were designed to see more of the outer aperture (e.g. f/1.4 lens), then it wouldn't work at all with f/2 or f/2.8 lenses (just as f/2.8 sensors don't work at all with f/4 or f/5.6 lenses). It only sees just that part of the lens that it's designed to (within certain tolerances... f/5.6 AF sensors will sometimes see a little of the f/8 ring, which is why they can kinda-sorta autofocus, though with a smaller baseline.)

[MikeWhy]

My understanding is that it's more than just a central occlusion: each of the paired AF sensors gets light from an aerial image of a virtual subaperture at opposite sides of the f/2.8 annular ring.
<div style="clear: both;"]</div>

I think I get it now. Instead of an actual ring, the paired sensors make up a portion of a virtual ring, the spacing being the same or at least analogous to the central occlusion. I'm quite sure I still don't understand why the ring or baseline distance improves accuracy (couldn't I just read adjacent pixels on the main sensor for contrast, for example?), but I already have a much clearer picture now of how the AF works. Thanks for the lucid explanations.

[neuroanatomist]

couldn't I just read adjacent pixels on the main sensor for contrast, for example?

Sure - in fact, that's exactly how contrast-detection AF works in LiveView on the newer bodies, and also how typical point-and-shoot cameras focus. It can be quite accurate - in fact, focusing a Canon dSLR in LiveView with a 10x magnification on the LCD is just about the most accurate way to focus a Canon camera. But if you've tried that, or used a P&amp;S camera, you know that relatively speaking, contrast detection AF is very slow, sometimes taking nearly a full second to achieve focus. That's one reason that drives P&amp;S users to dSLRs - it certainly was for me, too many missed shots of my fast-moving toddler!

[Daniel Browning]

You're welcome.

I still don't understand why the ring or baseline distance improves accuracy

Because the degree to which each paired AF sensor differs is directly proportional to the baseline distance. The further you get from the center of the aperture, the more that the light rays diverge. The f/5.6 AF sensors have a short baseline distance; their accuracy is 6 times worse than the f/2.8 AF sensors.