Canon EOS DSLR Autofocus Explained
by John Reilly (neuroanatomist)
With my first SLR, I placed my subject in the middle of the frame and turned the focus ring until a feature in the split prism lined up and the microprism collar lost it sparkly look. Today, things are a lot more complex and technological, and the number and variety of different terms used to characterize AF systems is a little bewildering. Hopefully, this article will clear up some of the confusion.
Phase detect AF
For this discussion, I'll be talking only about phase-detect AF - the rapid method used in dSLRs involving a separate AF sensor. The other method, contrast-detect AF, operates in Live View and is also found on P&S cameras. Although contrast-detect AF has some advantages, it is substantially slower than phase detect AF, which is the main reason the latter is the primary AF method for dSLRs. For phase detect AF, the main reflex mirror, which directs most light up to the viewfinder, passes some light through a semi-transparent window to a secondary mirror which directs the light down to the AF sensor at the bottom of the mirror box:
Before hitting the AF sensor, the light passes through an array of microlenses which separate incoming light to produce a pair of images that fall on each point in the AF sensor. A simple line sensor then measures the distance between those two images to determine if the image is front- or back-focused, and by how much. In essence, each AF point is operating as a simple rangefinder.
That's the simple case, but modern systems are much more complex - and there's a lot of jargon that make comparisons difficult. What do things like 'high precision' and 'dual-line zigzag arrangement' actually mean?
Characteristics of AF systems
The main characteristics of AF systems can be broadly classified as:
• Number of points
• Geometry of points - basic and complex
• Accuracy
• Precision
• Sensitivity
• Speed
• Coverage area
Number of points is self-explanatory. Geometry of points adds a layer of complexity, but both number and basic geometry are usually part of the top-line spec for an AF system (e.g. the T3i/600D has 9 AF points with 1 cross-type, and the 7D has 19 cross-type AF points). Complex geometry includes things like dual-cross points and dual-line zig zag arrangements. More on that later. Globally, accuracy is how 'close to true' the AF system gets, and Canon doesn't offer any specifications for this characteristic. Note that global accuracy is affected (and hopefully, corrected) by AF microadjustment. Locally, geometry affects accuracy, because a cross-type point has a better chance of achieving proper focus than a single-orientation point, since it samples features with different orientations.
Precision is repeatability - if you take several shots of the same thing, how close will the focus of any one shot be to all the other shots? For EOS AF points, there are two levels of precision - 'normal' which is within one depth of focus for the attached lens at max aperture, and 'high precision' which is within 1/3 of the depth of focus for the attached lens at max aperture. Note that depth of focus is related to, but not the same as, depth of field, although the same factors influence both.
Overall sensitivity is how well the AF system performs in low light. The AF sensor is composed of multiple 48-bit line sensors and associated amplifier circuitry - the more amplification (within the limits of signal to noise), the less light needed to focus. AF sensitivity is specified as an EV range, and the lower the first number, the better. For example, the T3i/600D through 5DII can AF down to -0.5 EV, the 1D IV/1DsIII down to -1 EV, and the new 1D X down to -2 EV. EV units are 'stops' so the 1D X can achieve AF in half as much light as the previous 1-series bodies. Sensitivity is used in another context as well, associated with lens aperture (e.g. an f/5.6-sensitive AF point) - in that case, it does not refer to the amount of light (sensitivity is perhaps not the ideal word).
Speed is another characteristic for which Canon offers no metrics, probably because there are too many variables. But better AF sensors and better processors result in faster AF performance.
Coverage area is a very important factor - the broader the area of the frame where there are AF points, the more likely that an AF point will fall on your subject. More on this later.
Of the above characteristics, number, sensitivity, speed, and coverage area are fairly straightforward. But geometry, accuracy, and precision are more complex, especially because the lens mounted to the camera will change the way in which the AF points operate in terms of geometry, accuracy, and precision.
Lines, crosses and double-crosses
A basic, single line AF point can detect contrast only in one dimension - the dimension 'opposite' to the orientation of the line. Remember that split prism in the manual focus SLR? The 'split' was horizontal, so you had to look for a vertical feature to focus on for the prism to be effective. So, a horizontally-oriented line sensor will detect vertical lines (like a flagpole or the side of a door frame), while a vertically-oriented line sensor will detect horizontal lines (like a horizon or a boat on the water). There's some confusing terminology here - a "vertical line sensor" is the same as a "horizontally-sensitive line sensor" and vice versa. A 'dual-line zig zag' arrangement (found on a few points in some xxD bodies and the 7D, and on all the points of the 1D X) means a line sensor that's actually two parallel lines instead of just a single line, and the pixels in those two lines are offset by half a pixel, meaning the point of maximum phase alignment can be more accurately determined because it will fall on a pixel in one line or the other, whereas with a single line it could fall between two pixels).
Generally, an aperture value is associated with an AF line sensor. The terminology usually used is "f/number-sensitive", e.g, you may have an f/5.6-sensitive line sensor, or an f/2.8-sensitive line sensor. The f/number refers to the maximum aperture of the lens, because AF is performed with the lens wide open (i.e. the aperture you choose for the shot does not matter, only the max aperture of the lens). The use of 'sensitivity' in this context implies that light levels matter, because that's what we think of when we normally use f/numbers. In this case, though, a wider aperture simply means a wider baseline for the rangefinder system is required for that line to function. Personally, I think better terminology might be to use threshold instead of sensitivity, so an f/2.8-threshold line would require an f/2.8 lens to function, and if you mounted an f/4 lens, that sensor line would not operate. An f/5.6-threshold sensor would work with any lens having a max aperture of f/5.6 or wider.
Note that these thresholds are not absolute - a lens with a narrower aperture than the threshold might still work, but at reduced effectiveness, accuracy, and speed. Thus, Canon limits the functionality to the rated aperture for a given AF sensor. However, some third party lenses (e.g. Tamron and Sigma zooms with a max aperture of f/6.3 at the long end) effectively trick the AF system into thinking there's an f/5.6 lens attached. Likewise, although not condoned by Canon, it is possible to use tape to block some of the contacts on a Canon 1.4x extender used with an f/5.6 lens, resulting in the camera attempting to autofocus with an f/8 lens on bodies which are limited to f/5.6. Sometimes, it even works...
All EOS bodies have f/5.6-sensitive sensors, and thus will work with any Canon EF or EF-S lens. Some 1-series bodies have an f/8-sensitive sensor at the center AF point, enabling them to autofocus (properly, and with official support) with an f/5.6 lens plus 1.4x extender or an f/4 lens plus 2x extender - a significant benefit for users of supertelephoto lenses. Notably, that feature is not included in the specification for the 1D X, which is limited to f/5.6 lenses for autofocus.
An f/2.8 sensor line is more accurate than an f/5.6 sensor line - the wider the aperture threshold, the wider the rangefinder baseline for triangulation and thus, the more accurate the measurement of focus. However, the wider the aperture, the fewer lenses that work with that aperture (and the more expensive those lenses are), and also, the detection range of f/2.8 sensors is narrower, meaning it may take longer for an f/2.8 line to achieve a focus lock when a subject is well out of focus. As a result, AF systems will usually focus in two steps when possible - 'coarse' focus with an f/5.6 line, then 'fine' focus with an f/2.8 line.
A cross-type sensor is a horizontal line sensor and a vertical line sensor at the same AF point, meaning that point is able to detect lines in both orientations - that makes it more likely that the AF point will be sampling a feature that has the correct orientation to activate the sensor. Some cross-type points have the same threshold for both sensor lines, e.g. the 9 points on a 50D/60D and all 19 points on a 7D are cross-type points with an f/5.6 threshold for both orientations. That's good, in that all the points operate as cross-type with any Canon lens, but the trade off is lower accuracy than f/2.8 lines. Some cross-type points only function as cross-type points when used with lenses that have a sufficiently wide maximum aperture, and if used with a lens with a narrower max aperture, they function as single line points only (usually vertical line sensors). This is a partial workaround to the trade-off mentioned above. A cross-type sensor with an f/2.8-sensitive horizontal line and an f/5.6-sensitive vertical line will function as a cross-type sensor with an f/2.8 or faster lens, but if you use a slower lens, you'll still have a functional AF point (but with only single-line orientation).
A dual-cross point is an even more complex AF sensor element. The cross-type sensors described above are a vertical line and a horizontal line, i.e.in the shape of a '+'. Dual cross-type points add another cross-type point in a diagonal orientation, i.e. an 'x' that is superimposed onto that '+'. All of the implementations of dual-cross points to date combine an f/2.8-sensitive diagonal 'x' with an f/5.6-sensitive orthogonal '+'. So, with a lens slower than f/2.8 you get a standard f/5.6 cross point, and with an f/2.8 or faster lens, you get the increased accuracy of an f/2.8 baseline, and with the ability to detect lines in multiple orientations.
Precision, invisibility, and other intangibles
As mentioned above with the precision discussion, there's a modified type of AF point called a 'high-precision' point, which focuses within 1/3 of the depth of focus of the lens at max aperture, vs. the normal precision spec of within 1 depth of focus. Usually, the high precision point is the center point, and it operates in high-precision mode with an f/2.8 lens on most bodies, or an f/4 lens on 1-series bodies. The 1D X is an exception in two ways - it has five high-precision points in a central vertical column, instead of just one, and they require f/2.8 unlike previous 1-series bodies.
On some 1-series EOS bodies, certain AF points are 'assist' points, meaning that they cannot be selected when manually choosing a focus point. However, they will be automatically selected by the AF system, either when adjacent to a selected AF point with an AF expansion mode enabled in One Shot mode, or when needed for tracking a moving subject in AI Servo AF mode. The 9-point AF system found in the 5D and 5DII also has 6 assist points, but unlike the 1-series assist points, these are 'invisible.' They are used in AI Servo AF mode to help with subject tracking, but are not available at all in One Shot mode.
Another feature worth mentioning is extreme defocus detection. When nothing is even close to correct focus, it's a challenge to the AF system, which mostly acts by refining an image that's somewhat out of focus. Older AF systems simply racked the focus back and forth until some feature(s) came into close enough focus to activate the sensor lines. Newer sensors employ 'extreme defocus' sensors to point the AF system in the right direction with a preliminary estimate of the direction and magnitude of the change needed to get close to correct focus. These are actually the same sensors as the 'dual-line zig zag' sensors described above - that configuration increases accuracy, but also allows better defocus detection by 'summing' the offset lines into a single readout. Importantly, that extreme defocus detection uses combined data from the dual-line zig zag sensors. On xxD bodies and the 7D, three vertical points have the dual line vertical sensors (center top, center, and center bottom), and data from those three inputs drives the extreme defocus detection. On the 1D X, all of the vertical lines (i.e., all 61 AF points) have the dual line zig zag arrangement, and thus the whole AF sensor acts as one big extreme defocus detector.
Separate from the number and characteristics of individual AF points is the ability of those points to work together to track a moving subject in AI Servo AF mode. Because f/5.6-sensitive points generally acquire focus faster than f/2.8-sensitive points, the AF system will often rely on them for tracking data. However, overall tracking performance depends on many factors outside of the AF sensor itself, most notably the Digic processor and the algorithms driving the motion prediction, as well as supporting data integrated from other sources (e.g. the metering sensor on some EOS bodies).
Putting it all together
That's a lot of features, and many of them are combined in various ways to compose an AF sensor. Here's an example how it all comes together, in the AF sensor layout of the 7D:
Note that this is the entire AF sensor, even though at first glance it appears to be just one large cluster with two smaller clusters at the sides, such that you might be tempted to think it's the center AF point and the two on either side. But in fact, you're looking at 20 AF points (19 '+' f/5.6 points and one 'x' f/2.8 point superimposed on the center '+' point). The way to interpret this diagram is that each adjacent pair of lines in a horizontal orientation, e.g. — —, represents the horizontal sensor lines of one to three AF points (longer lines contribute to more AF points), each adjacent pair of lines in a vertical orientation represents the vertical lines of one to five AF points (longer lines contribute to more AF points), and the pairs of single diagonal lines represent the f/2.8 'x' cross-type point.
Here is the actual AF sensor for the 1D X. The set of five f/2.8 diagonal crosses stands out pretty clearly.
When looking at the diagonal f/2.8 sensors, it's apparent that they are much more widely spaced than the f/5.6 sensors - almost to the edges of the chip. This accounts for the longer baseline that results in greater accuracy than the f/5.6 sensors.
AF Point Coverage
For many people, this is a big issue in comparing cameras. While it would be wonderful to have AF points available over the entire extent of the frame, There are technical limitations on the spread of the AF points - at best, they can only occupy the middle area of the frame, because of simple geometry and optics. In a nutshell, there are four reasons for this limitation:
- Size of the secondary mirror. Light for AF passes through the semi-transparent part of the main mirror (most is reflected up to the viewfinder), then is reflected off the secondary mirror down to the AF sensor. There is limited space behind the main mirror, based on the necessary geometry (i.e. the main mirror has to be at a 45° angle to the incoming light, and the secondary mirror has to be behind the main mirror and at an angle of 90° to the main mirror, so it's length is limited by the distance between the main mirror and the image sensor).
- Distortion. With many lenses, the edges of the frame are subject to distortion (barrel/pincushion), and that reduces the accuracy of phase detect AF.
- Vingetting. The AF system needs a certain amount of light to work. Almost all lenses vignette to some degree, meaning there might not be enough light at the edges of the frame. For example, the EF 17-40mm f/4L has >2 stops of vignetting wide open at the wide end - that means at the edges of the frame, you're below f/5.6 and AF sensors would not have enough light to operate (i.e. in dim light you'd be below the EV sensitivity of the sensor).
- Temperature. Canon has stated that larger AF sensors are more susceptible to changes in temperature with the result that they change size, getting either larger or smaller as the temperature rises and falls. That reduces the accuracy of the AF system overall.
It's worth noting that none of these limitations apply to contrast detect AF, so using LiveView you can autofocus right out to the edge of the frame.
So, AF sensors are limited to the central portion of the frame, but some bodies offer more extensive coverage than others. A common question is, "What does the relative spread of AF points look like?" This image comparison below shows the AF point coverage of recent xD bodies (7D, 5DII, 1DsIII, 1D IV, and 1D X):
The 1D IV and 7D are about tied for the broadest spread in both dimensions (relative to frame size). The 1D X has the same lateral spread as the 1D IV and 7D, and slightly less spread in the vertical dimension (about half a row shorter). The 5DII is in blue, and you can see the narrow spread relative to the other xD bodies (the horizontal extent of the 5DII's AF points is the same as the 1DsIII, but the vertical extent is much less).
In addition to the area covered, density of points is also important. A more densely packed array of AF points will gives more options for selecting a focus point. More importantly, a more densely packed array delivers substantial improvements in AI Servo AF tracking, because subjects pass from one point to the next more quickly - less lag and more data to drive the predictive algorithms.
Final thoughts
Hopefully, the above information is helpful in evaluating the characteristics and relative strengths of Canon EOS autofocus systems. Although the AF system is only one feature of many to consider when comparing camera bodies, for many people it's an important one. Historically, Canon has differentiated the levels within the EOS lineup partly by using different AF systems, with AF performance and particularly AI Servo AF tracking getting progressively better as you move from the xxxxD bodies up to the 1-series (with the 7D as an outlier, with the best AF system outside of a 1-series body). Keep in mind that just like image quality is about more than megapixels, AF performance is about more than the number of AF points.
--John
