Autofocus

Continuous AF: Closer Distance Subject Tracking

Cameras are increasingly getting better at automatically choosing focus points to stay on subjects as they move around the frame, an ability we refer to as 'subject tracking'. And as we saw on our last page, the 5DS/R generally subject tracks relatively well when it comes to long distance subjects well-isolated in depth (a typical sports scenario), although it can lag behind the subject during continuous shooting. The depth isolation aids systems that predominantly use depth information to subject track, which has traditionally described Canon systems, while the long distance tends to make things easier on the focus system, since a moderate change in distance for a distant subject requires only a relatively subtle refocus.

"Cameras are increasingly getting better at automatically choosing focus points to stay on subjects as they move around the frame."

However, systems that rely heavily on depth information (e.g. 5D Mark III) can struggle significantly with close-up work, where there may be many different subjects within a similar depth; an eye vs. a nose, for example. At these distances, for these subjects, image analysis and scene understanding yields better results, as these methods can distinguish features from other distracting features in the scene. See our demonstration of the advantages of subject tracking using scene analysis over depth information only.

The introduction of iTR to the 5D line, enabled by the addition of a 150,000-pixel RGB metering sensor, should bring scene analysis to the AF system, theoretically aiding it to understand and track subjects. The RGB metering sensor is essentially a low resolution image sensor that 'sees' the scene at all times. So we put it to the test, trying to track the eye of our mannequin. This test, below, simulates close-up portrait or candid photography with moving subjects, with our movement simulating both recomposition and subject movement.

Here Canon's iTR system has a lesser degree of success, getting easily confused and losing track of its original target, typically the eye of our mannequin. The performance improves with a well isolated subject, such as our cyclist on our previous page, and we think that this overall discrepancy is due to the system relying very heavily on depth information for subject tracking, rather than the pattern and color information from the camera's metering sensor. In other words, the system appears to function largely like previous Canon DSLRs without a metering sensor, primarily using depth information to subject track.

"Subject tracking with iTR doesn't work as well as the camera can perform for subjects at greater distances with telephoto lenses... nor as well as Nikon's 3D tracking or Sony's eye tracking."

This unfortunately simply isn't great performance: subject tracking with iTR doesn't work as well as the camera can perform for subjects at greater distances with telephoto lenses. Nor does it work nearly as well as the Nikon D810's 3D tracking, or the Sony a7R II's eye tracking, both of which are very good at keeping track of close-up subjects using image and pattern analysis for subject recognition and tracking.

Real-World iTR Tracking of a Toddler

We put iTR to the test in a real-world scenario to see how the poor performance of iTR in our mannequin test above correlates with real-world performance. The camera was set to use any of its 41 cross-type AF points to follow the subject we initiated focus on with the center point: in this case, the eye of our toddler. Ideally, the camera would automatically choose the appropriate AF point(s) to stick to our toddler's eye as she moves or as the image is recomposed. Such subject tracking is not just useful for shooting toddlers: candid portraiture and wedding photography can benefit from the ability to track an eye or face so that it's in focus at the decisive moment of capture, even if your subject moves, throws his or her head back in laughter, or you recompose.

As you can see below (click on the large image for full resolution), iTR often doesn't successfully track our subject's eye, which was placed under the selected AF point when focus was initiated. It gets easily confused by other features on her face. The result is that many shots are completely misfocused, with a hit-rate of approximately 30% with a 35L shot at F1.8. The general pattern of sticking somewhere on our toddler's face, but not specifically on the eye, corroborates our hypothesis that the AF system is primarily using depth information, instead of subject and pattern recognition, to subject track. Such a system tries to choose AF points registering a distance reading similar to the initial distance reading (accounting for approaching or receding motion), but cannot hope to be as precise as a system relying more heavily on actual image analysis to understand a subject and track it.

For example, take a look at the nearly perfect hit rates demonstrated by the Sony a7R II in a similar test with a Sony FE 35mm F1.4 lens, as well as the D810 in 3D tracking mode.

Image 1
Image 10
Image 2
Image 9
Image 3
Image 8
Image 4
Image 7
Image 5
Image 6

Real-World 'Expand AF Area: Surround' Performance

Since iTR was largely unusable to track our toddler, we reverted to a more traditional manner of shooting: using Canon's 'Expand AF Area: Surround' mode, which prioritizes the AF point you've chosen but allows the camera's tracking system to use any of the surrounding 8 points to focus on the subject if the camera thinks your initial subject is now under one of them instead of the chosen AF point. This can be useful for erratic toddlers, since keeping a single, rather small AF point over the eye of a moving toddler can be relatively difficult. See the performance of this mode below (click on the larger image for full resolution).

The hit-rate is around 50%, only a little better than what we saw with iTR, but far below what we expect using a single AF point with this system. It appears that any time you give the camera some freedom in selecting its AF point, it gets easily confused, either selecting the wrong target (as we saw above), or simply missing focus entirely despite our constant reframing to keep the selected AF point over our subject's eye in each shot, as we see below. For reference, a Nikon D810 paired with a similar 35mm prime had nearly a perfect hit rate in this same test.

Image 1
Image 16
Image 2
Image 15
Image 3
Image 14
Image 4
Image 13
Image 5
Image 12
Image 6
Image 11
Image 7
Image 10
Image 8
Image 9