Technical Discussions

Autofocus Reality Part 3B: Canon Cameras

Published August 1, 2012

Some days it’s good to be a geeky gear-head. This is one of them because a) I finished testing autofocus on all of the Canon camera bodies we had, b) I actually found out some interesting stuff, and c) I got worked up about camera marketers while doing it, so I have my next post in the works already.

Because I’ve been writing this series on the fly (telling you what we found as we found it), I’ll keep going in that fashion and keep the conclusions for the end of the article.

What We’ve Found So Far

In the first Autofocus Reality article, we demonstrated two things:

1. Phase-detection autofocus (even using still targets and center-point only) wasn’t nearly as accurate as contrast detection.

2. The contrast-detection autofocus was about as accurate as the most careful manual focusing.

Part two of the series showed that a few newer lenses did focus as accurately as contrast detection on 5D Mark III cameras but not on 5D Mark II cameras. The third article (part 3A) showed that the newest Canon lenses (40mm f/2.8, 24mm f/2.8 IS, 28mm f/2.8 IS, 70-300mm L IS and 300mm f/2.8 IS II) focus more accurately when mounted to 5D Mark III camera but not on 5D Mark II cameras.

The Canon 70-200 f2.8 IS II, oddly enough, seemed not quite as good as the above lenses, but more accurate than the older ones. Whether this was oddness in my measurements or a real finding, I wasn’t sure.

Since we knew that the newest lenses autofocused accurately on 5DIII but not on 5DII cameras, the next step was obviously to compare an accurate AF lens on different camera bodies to see with which ones it was capable of accuracy.

We expected the 1Dx (which has the same AF system as the 5DIII) would be accurate. We weren’t sure about the others.

Today’s Contestants on The Focus is Right!

We know the Canon 28mm f/2.8 IS II had accurate autofocus on the Canon 5DIII so we chose one copy to be our test lens.

We AF microadjusted each camera to the lens prior to shooting. If the camera did not offer AF microadjustment, we checked the body with the test lens and exchanged it for another copy if the lens back or frontfocused at the test distance.

(Microfocus adjustment makes absolutely no difference in the shot-to-shot variation in AF–it only improves the average value of the group. But I got tired of explaining that to people in the previous articles. It was easier to just do it than to answer another 50 emails.)

We then tested it in our Imatest lab using one copy of each Canon camera we carry. To save you from running amok on the internet, finding out which cameras are how old and have what kind of autofocus, I’ve listed that information in the table below.

 

Camera Year released Year firmware AF description
1DsIII Dec-07 12/16/09 45 point, 19 cross, 26 assist, f/4 at center, dedicated AF processor
5D II Dec-08 2/28/12 9 point, f/2.8 cross center
50D Dec-08 2/28/12 9 point cross, dual diagonal center cross
7D Aug-09 4/25/11 19 point, all cross, center dual diagonal cross (advanced algorithm), first “zone AF” and “spot AF”
1DIV Dec-09 3/29/12 45 point, 39 cross (f/2.8-f/5.6, f4 at center), first AF expansion camera
60D Aug-10 6/19/12 9 point, f/2.8 cross center
T3i Jan-11 1/30/12 9 point, f/2.8 cross center
1Dx Mar-12 2012 61 point reticular, 41 cross type, 5 dual diagonal cross, including center
5D III Mar-12 2012 61 point reticular, 41 cross type, 5 dual diagonal cross, including center
T4i Apr-12 2012 9 point, f/2.8 dual cross center, hybrid CMOS AF Live View

 

I suspected that the autofocus improvement we’ve seen had more to do with hardware than firmware, but I listed both year of release and year of latest firmware upgrade for completeness. I also listed the basics of the camera’s AF system, as well as any marketing comments made about it at release like the 7D having “the most advanced AF algorithms.”

My thought going in was the difference would be in hardware not firmware. I don’t think any amount of firmware is going to make the AF sensor on the lower left behave like the one on the lower right, even in center-point, single-shot mode.

 

The Results

In the lens variation article, we used the standard deviation (SD) to measure how accurately the camera focused: Softer focus results in lower Imatest values. As an example, we’ve put up a graph of Imatest values for the 5D Mk II versus the 5D Mk III below.

 As you can see, the 5D Mk III shots (red square) are all very similar. The 5D Mk II shots (blue diamonds) are more spread out: The variation in shot-to-shot focus is greater.

In this example, the SD of the 5D Mk III samples was 17 lp/ih, while the SD of the 5D Mk II was 38.5. Those are similar to the numbers we’ve seen over and over—accurate focusing combinations have SDs in the teens while less accurate ones have SDs in the 30s.

Rather than clog up the post with a lot more graphs, I’ll list the SDs of the various cameras with the 28mm f/2.8 IS lens in the table below.

Camera SD
1DsIII 29
5D II 38.5
50D 34
7D 41
1DIV 22
60D 34
T3i 41
1Dx 17
5D III 17
T4i 29

 

It’s a little confusing. There’s a range of variation, of course. But clearly the 5DIII and 1Dx do better than the other cameras, while the 1D Mk IV seems to be a bit between those two and the rest of the pack.

It seems a little clearer to me if we graph the standard deviations and separate the cameras by type (more expensive at the top, less expensive at the bottom.) The oldest cameras in both groups are on the left, while the newest are on the right.

 

A couple of points are worth making.

The graph of the more expensive cameras seems to show a pretty logical progression. The 1Ds III is by far the oldest, having been around since 2007. But it included every possible AF technology of the day, including a separate AF processing chip. The 5D Mk II, even on its release, was known to have “consumer-grade” autofocus.

Despite my well-recognized modesty, I will also point out that when the 5D Mk III was first released, and Canon fanboys were dropping off cliffs right and left, I said “the 5D III is no minor-upgrade camera; it’s an entirely new camera using the old camera’s name”. Its autofocus system is certainly not a minor upgrade–it’s moved over to the big-boy camera side.

I had hoped the T4i might be more accurate than it was, at least with new lenses. It does seem more accurate than the other consumer / prosumer cameras in phase detection, but it’s not nearly as good as the 1Dx or 5D III.

I assume that it’s new hybrid LiveView system does not carry over to create phase-detection AF. I will say, in it’s defense, that when focusing in LiveView it is obviously faster than and just as accurate as any of the other cameras, including the 5DIII and 1Dx.

So Why Could This Be?

All this autofocus stuff 1) gave me a headache and 2) made me rather curious and uncertain.

I started doing a simple demonstration of what I already knew: Phase-detection AF isn’t as accurate as contrast-detection AF. But then I got results that indicated sometimes it is just as accurate as LiveView. But you have to have a certain camera and a certain lens or it doesn’t happen.

This didn’t make much sense to me.

I would have understood if each generation of newer cameras and lenses got a little better. Or if a new camera or lens was dramatically better. But why a rather sudden change, and why did you need both a new camera and a new lens?

I spent a fair amount of time emailing with Dave Etchells of The Imaging Resource  and SLRGear.com who was kind enough to bounce ideas back and forth with me. His thoughts sent me on the right track for figuring this out.

I started off by trying to find out how long it took for an autofocus improvement to go from idea to released-to-the-public. I found one answer in a Canon patent from 2003 describing combined phase / contrast-detection AF in the camera mated to a lens using a stepper motor.

Which we saw  . . .  oh, yeah. Now. With the Canon T4i / EOS-M sensors and new STM motored lenses.

One thing of interest in this patent application, tucked away in the background section, is the following statement:

. . . though the prior art indicates a method for realizing high speed and high precision of autofocusing at the same time, it does not accompany a lens drive control for realizing this and thus does not adequately realize [. . .] high precision of autofocusing.

In other words, back in 2003, Canon recognized that a high-accuracy AF system in the camera required a more accurate lens drive to yield precise autofocus. You can’t have one without the other.

Suddenly the need for both a new camera and a new lens to get accurate autofocus began to make sense. But wait, the patent was talking about using a stepper motor to achieve accurate AF. We did find the 40mm pancake (stepper motor) was more accurate. But we also found that the new 24mm, 28mm f/2.8 IS and 300mm f/2.8 IS II lenses were more accurate. Yet they don’t have stepper motors.

Then I read on a bit further in the patent:

In order to achieve this objective, this invention provides a camera system comprising: a first focus detection unit, a second focus detection unit, a stepping motor that drives a focusing lens, . . . or a rotation detector, which detects the rotation . . . of the motor. . . The control circuit performs closed-loop control, based on the output of the rotation detector to control the motor.

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.

This is backed up by a lot of other information, including statements in Canon’s “EF Lens Work III” and quotes made by a lot of people who would have known.

It’s clear that later AF systems were closed loop (the camera double checked where the lens had gone), although exactly when and where that change occurred I don’t know. But this probably explains the lengthy, and now largely meaningless, debates about whether systems were open or closed-loop: They used to be open. Now some (probably most) are closed.

More to the point, though, is the comment that a rotation detector would be needed in lenses driven by ultrasonic motors to increase accuracy of the focusing movement.

If this is the case, then the newer Canon lenses should definitely have a rotation detector built into them. We know there are rotation detectors in many lenses released after 2000, but if they are  in older lenses we can’t identify them, so this fits too. (As an aside, I am particularly skilled in finding them because usually if you touch them with your fingers the lens won’t focus anymore and the unit has to be replaced.)

 

From Left: A Canon and Nikon magnetic rotation detector.

 

Just to be certain, I asked Aaron to find an excuse to take apart one of the new lenses. Or failing that, I told Aaron to take one apart and not tell Tyler or Drew, who never really believe we can put them back together correctly.

The result: Aaron checked and the newer lenses did have rotation detectors.

There was one other big hint hiding in plain sight in Canon’s resource article on precision cross-type AF sensors. The article states that the farther apart each pair of AF sensors are (each pair compares phase from opposite sides of the lens) the more accurate the sensor. It also states only the 5D Mk III and 1Dx have them. I italicized some of the quote for emphasis:

[…]simply by reading which pixels on each sensor line are being struck by light, the AF system can tell instantly what direction to move the lens in for proper focus, and by how much to move it. . . Canon EOS SLRs with high-precision AF sensors simply move the pairs of sensors much farther apart, and accordingly, the AF information can be more finely broken-down and reacted to.

The 5DIII and 1DX have f/5.6 cross-type sensors (most previous cross-type required f/2.8 or faster lenses). The central AF sensors in these cameras also have diagonal cross sensors. The article continues:

For pros who prefer to use the center AF point, simply manually selecting it [. . . ] gives the user the highest precision AF possible with these cameras (assuming an f/2.8 or faster lens is being used). And, with their unique diagonal cross-type layout, the AF points are much more likely to latch-on to typical horizontal or vertical subject details.

In Summary

As it turns out, my little tests just confirm things that were already published (if obscurely.)

The two newest Canon cameras have more accurate phase-detection sensors than their previous cameras. The newest lenses have more accurate focus movement (or provide more accurate focus movement feedback, or both) that takes advantage of those sensors.

Older cameras don’t have accurate enough AF sensors to take advantage of the new lenses’ capabilities.  Older lenses can’t move their focusing elements with enough accuracy to take advantage of the new cameras’ accurate sensors.

It’s rather sad (given the amount of other work that’s piled up during this little series) that I didn’t find this information until after I’d done all this testing, but it’s rather a testament to our times. There’s so much marketing drivel and useless verbiage thrown about that the marketing noise drowns out the actual useful information the camera makers offer us.

But that’s the topic of another blog post in which I will take personal offense at the marketing crap that’s been shoveled our way at an ever-increasing rate.

I don’t believe we photographers are nearly as stupid as the marketers seem to think we are. In the meantime, while I’m on this rant, I highly recommend Thom Hogan’s very funny look at camera icons “A Different Kind of Focus.” Thom writes better stuff sitting on the tarmac waiting for takeoff than I do at my desk.

And I guess, given all the marketing fluff, I wouldn’t have really believed “better autofocus” if I hadn’t seen it with my own tests.

I’ve heard it before and it wasn’t so. Not to mention I doubt seriously we’d have seen the marketers say, “better autofocus…but not with the lenses you already have.” So maybe this is worthwhile after all.

Otherwise I might have put my trusty old 85mm f/1.8 on a 5D III and thought “not better at all, they’re lying to me again.”

 

Roger Cicala (with thanks to Aaron Closz and Dave Etchells)

Lensrentals.com

August 2012

Author: Roger Cicala

I’m Roger and I am the founder of Lensrentals.com. Hailed as one of the optic nerds here, I enjoy shooting collimated light through 30X microscope objectives in my spare time. When I do take real pictures I like using something different: a Medium format, or Pentax K1, or a Sony RX1R.

Posted in Technical Discussions
  • K D Sandmann

    Thank you for yet another great article.

    To some extent it seems people using older glass are now faced with updating the glass to justify & fully benefit from buying new bodies.

    It would be nice if Canon would give us a list of optimal combinations but I doubt that would help their sales. O well. 🙂

  • Robbie

    The older AF lenses definitely have a rotational position sensor. It may not be as precise as the newer ones but it exists. I have seen it inside a Canon 50 1.4

    It looks like the below image, with brushes riding against traces along the barrel (measuring resistance that varies as it is turned, ie a potentiometer).
    http://img18.imageshack.us/img18/2586/img9645x.jpg

    The EOS control scheme must be closed loop on all the lenses… when the lens is installed or switched from MF it knows where the focus is rotationally. Otherwise it would need to drive to MFD or infinity focus first to establish a known starting position, which from experience doesn’t occur. (Similarly for focus distance for flash in MF mode.)

    It is likely the newer lenses have sensors with more angular resolution.

  • RP

    Great job Roger. I always realized my 7D, and my 5DMKII too, were not so precise to auto-focus. I think inter shot variation should be software rather than hardware dependent. The only thing I was surprised was to know that T4i does not focus so precise as we should expect from the acquisition of the new sensor technology!

  • Good job! Thanks!

  • Interesting stuff, great to see rigorous testing in the AF area.

    It’s important to identify which control process you’re talking about. The Ishikawa patent is about the control system _within_the_lens_, which interprets commands from the control system _within_the_camera_. It’s the second system which has been debated at length and proven to be closed-loop from at least the 20D onward (see http://www.dpreview.com/articles/5402438893/busted-the-myth-of-open-loop-phase-detection-autofocus).

    Whether the control system in the lens is open or closed-loop doesn’t make any difference to the question of whether the control system in the body is open or closed-loop, and it’s invalid to make conclusions about the second from an old patent about the first.

  • Ethan

    It’s not totally clear to me how much of a difference we’re talking about here. Is the variance noticeable shot-to-shot? How much of it is related to your focusing methodology (spin ring, refocus – I know you mentioned better results just by hammering the autofocus button a few times in a row)? If I have a 5DMIII, is it worth waiting for a version III of the f2.8 70-200 w/ IS or is that just silly?

  • Zak

    oh, oooh!
    One more thing! I’d love to see whether 3rd party lenses can also do this trick (but suspect it will take them some time to do so, if they manage at all). Even if they can’t: Is there a significant difference in the phase detect AF scatter between Canon and third party lenses?

  • Zak

    First: Thanks a lot for (not just) this very enlightening series of posts. I know no camera or lens review site whose operators could not learn a lot from your systematic approach to pretty much anything.

    Second: I’d imagine that the standard deviation values are difficult to compare across different cameras because their sensors have different resolutions. If the lens manages to outperform the sensor, resolution-wise, there’s some space for the focus motor to move without imatest reporting much (or at least not as much as otherwise). I’d also be interested in knowing if the resolution figure is actually properly linear with the focus error.
    Not that I want to cause you more work, but it would seem that the proper way of testing which camera supports the feature would be to compare two lenses (an old one, a new one) with similar optical properties (i.e. best achievable center sharpness) on each camera and see which bodies will show a difference. That way you can separate the difference in the autofocus procedure from the difference in the sensors. … having said that, it might also be overkill 🙂

    Another data figure I’d be interested in: Is the new autofocus procedure faster or slower than the old one, or neither?

    My background is in science and engineering (including optimization), and the computationally cheapest algorithm for minimizing a function is the Newton method, where you have the value of the function (offset on the AF sensor) and the gradient (how much movement gives me which change in the offset?), or some approximation of it. You compute how fare you need to go, then re-evaluate and so on until you’re satisfied. As you’ve shown, most cameras seem not to re-evaluate too much in favour of being done faster (or their AF-sensors are not that accurate?). But if you can re-evaluate on the fly, while the lens is still moving, and if you get feedback from the lens about how far it moved in reaction to the last input and what that did to the AF sensor offset — you have more data on your hands and can move from first-order approximation (Newton) to second-order or even some much fancier algorithms to know much more precisely how to correlate input and output.

  • Quick question. I suspect your manual focus tests were done using either live view, or the standard Canon-installed focus screen…which I find to be pathetically poor for manual focusing. Do you think manual focusing would improve significantly, using an after market focus screen such as that supplied by KatzEye? I’m a 7D user. The work you folks put into these articles is greatly appreciated!

  • CarVac

    I thought that most ring-USM lenses have at least a coarse sort of distance sensor for flash use, but I guess that’s not the same as this more precise sensor for autofocus feedback.

  • Tern

    “The 5DIII and 1DX have f/5.6 cross-type sensors (most previous cross-type required f/2.8 or faster lenses). ”

    Any chance you’d test other brands, such as Nikon/Pentax whom have had f5.6 cross-type sensors for generations now, and most interestingly with the new f8 cross-type sensors in the D800?

    It’d be interesting to see if the theoretical choice of ‘accuracy (if you have f2.8)’ (Canon) vs ‘availability (cross point with any lens)’ (Nikon/Pentax) actually matches to reality.

  • Roger Cicala

    Hi Shane,

    I do not have such a list. We only happen to know if we’ve looked inside and seen it. And it’s possible for us to look inside and miss it if we haven’t completely disassembled the lens.

  • Michael B

    Thanks for the insights and your article!

    I always wondered why my 40D needs 2 or 3 AF actions (manually induced by half-pressing the shutter butten) before further AF trials gave the same result (e.g. with f/2.0 100). Since I observed this I trigger the AF several times until I do not hear any noises from the USM drives.

    I think that I induced something like a feedback loop manually (by changing my brains “software” for photographing) – perhaps this might be an idea to recheck the older cameras and compare the AF consistency with that of the newer ones.

    Best – Michael

  • Mel Gross

    Oops! Sorry for the typo’s. I’m typing on my iPad which is on my lap, in a moving vehicle, and didn’t notice the errors until
    posted.

  • Mel Gross

    This is interesting, As it follows CNC machine tool practice, something that all camera and lens manufacturers are very familiar with.

    There are two types of CNc machines. The least expensive ones use stepper motors to move the table, quill, or forth or fitth motion. These are accurate enough for those machi ES, and can drive 0.001″ accuracy. On very well made machines, it can even go down to 0.0005″. But that where it ends. And today, a half thou. Isn’t all that great. But this is an open process, no checking of where things are.

    Better machines use servo motors. As can be understood from the name, these motors use a feedback loop going back to the computer, which knows where they are at all times. These machines are also made more precisely. They can have an accuracy to 0.000005″, or even twice that—one hundred thou. Inch! That’s accurate! A compromise has arisin more recently that uses steppers with feedback, and is I between the two, though closer to traditional steppers. Still, it can get to a “tenth”, as we call it, which means a ten thousandth.

    So what Canon is saying here makes great sense, but requires a different way of thinking, and a higher level of manufacturing reliability, something we’ve seen hit within the mid, 90’s. It’s too bad that it takes so long, but figuring out how to do this without breaking the bank isn’t easy. In addition, sensor making has hit a new high during the middle of the last decade, which has allowed this from the other end.

    It’s good to see it in action in cameras, though too bad that all of our lenses will need replacing (assuming that the mis focussing really has any effect most of the time in the real world of shooting.).

  • very interesting and informative Roger well done!!!!

    Roger do you happen to have a list of Canon lens that have the Rotation detector or can you point us to a list if one is already published else where?

    warm regards

    Shane

  • Great article! I’m sure the engineers from Nikon, Sony et al. are just as interested in this as we photographers are.

  • Roger Cicala

    Hi Jon,

    I’ll certainly do NEX when I try to test mirrorless, but I don’t have a good comfort level with alpha autofocus. I’m afraid I might make technical errors since I’ve never done this with SLT mirror cameras.

  • Jon

    Any hope for a Sony test?

  • hwyhobo

    Good stuff, Roger! Having had to make sense out of technical publications for most of my life (I work in technical training, so I actually have to understand it, not just memorize keywords), it is a pleasure to read a straight-shooting article.

  • Will

    Thank you very much for this very interesting review . I keep my eyes on your future stuff on Nikon and mirorless gears !

  • Jos

    …It’s a little confusing. There’s a range of variation, of course. But clearly the 5DII and 1Dx do better than the other cameras, while the 1D Mk IV seems to be a bit between those two and the rest of the pack…

    5DII must be 5DIII

  • Roger Cicala

    Mark, I couldn’t agree more. My favorite example, now mercifully pulled off of their website, was a Tamron blurb for an f/5.6 Vibration Control lens that said “Vibration Control allows you to freeze motion, making it perfect for athletic events”, or something very similar.

  • Maji

    Great series Roger. You are not only technically gifted (an euphemism for techno geek), but you write well too. Not many techno geeks can claim that. You maybe in the cross hair of the marketing consultants as you are trying to cull out their babble. Keep this up… I mean keeping us educated and informed photographers 🙂

  • markb3699

    As a former marketing writer (now a full time photographer) I’d like to point out that the reason for all the marketing hype may be due to the fact that the people writing Canon’s marketing content may simply not fully understand the products. I don’t think many product managers don’t understand product lines they’re responsible for with great depth. It didn’t bother some of the large companies that I worked for that I didn’t ucomprehend their products in fine detail, only that I could explain them to a reasonable degree and make brochures and data sheets read smoothly.

  • Roger Cicala

    Andre I’ll do Nikon next, and then go on to mirrorless.

  • Andre

    Roger, this is really fantastic stuff — thank you! Any plans to check on the accuracy of mirrorless cameras’ CDAF?

  • Roger Cicala

    Doug,

    I wish I understood the 7D system better. I’m not sure if it was an experimental dead end, or maybe is improved and just not in ways that this kind of test (center point AF on still subjects) detects. Certainly what I read from 7D users seems to show that there’s a learning curve with it’s AF system, but otherwise opinions are pretty wide spread. A few people love its AF, a lot don’t. I expect that means it’s good for certain situations more than others, but I don’t know what those are.

    Roger

  • Maiaibing

    Fits my experience with the 300 f/2.8 IS L I, 35L and 135L which do not seem to focus notably better with the 5Diii than the 5Dii.

  • Doug

    Great series of articles Roger, as always! I always enjoy reading your work as you are so thorough with every aspect of what you are doing.
    So, I’m noticing that the 7D is tied for the highest standard deviation in your Imatest values chart – when you say “Older cameras don’t have accurate enough AF sensors to take advantage of the new lenses’ capabilities. Older lenses can’t move their focusing elements with enough accuracy to take advantage of the new cameras’ accurate sensors” does that include the 7D as one of the ‘older cameras’? And if so, what does this mean for 7D users – is it worth upgrading/using any of the newer lenses as a result of this focus inaccuracy? Or are we still going to be far better off with the improved glass, despite the inability to take advantage of the newer focus capabilities?

Follow on Feedly