I thought I would weigh in with my experiments and experiences comparing the AF on the Canon 5Ds and the Canon R6. I apologize for the length, and there is a tl;dr at the end.
I'm going to divide AF into two functions: focus tracking which is distance to subject under a selected AF point, and subject tracking which is selecting the AF point over the subject. Some of the 'why' is subjecture on my part, so take it with a grain of salt. An actual camera engineer might come by and tell me I'm wrong on those points.
5Ds AF Design: The 5Ds AF is derived from the 1DX with an improved 150K pixel metering sensor that provides subject recognition (iTR tracking). 61 pts, 41 cross type, 5 double cross type vertical center. There are a couple differences versus the 1DX. In single shot mode the 5Ds will take a little longer to verify and adjust focus since the 50mp sensor will reveal very slight focus 'errors' not visible on an 18mp sensor. Sometimes this includes a secondary lens movement. I do not believe it does this in continuous AF, or at least I have never seen this delay in that mode. If I remember correctly, the 1DX can drive a lens faster on initial acquisation (literally more power to the lens motor), and I do not believe the 5Ds has this ability. There may be other differences such as processing speed or PDAF sensor readout frequency.
R6 AF Design: The R6 has DPAF II which IMHO is the first time Canon mirrorless cameras could compete with upper tier DSLRs on focus tracking. The entire sensor is a massive AF array, and it uses the sensor for both focus and subject tracking. One weakness is that there are no cross type points, which the article mentioned. The camera cannot AF on detail which is perfectly aligned with the AF points and you have to tilt the camera slightly for it to lock.
Focus Tracking: purpose built DSLR AF sensors are very, very good at this, and appear to still be competitive with mirrorless. Using a selected AF point in continuous mode the 5Ds can lock and track subjects as well as the R6 and sometimes even better or faster. A practical example of this: at an airshow I will put the center AF point on a plane when it is still a dot in the distance so that I'm smoothly following it as it approaches flight line. The 5Ds will quickly and confidently lock and hold that dot. With the R6 I found that I had to wait until the plane was a little bit larger in the VF to hit the AF button. If it's too small the R6 may start to hunt, and is slower on initial lock even if it doesn't hunt. (When the plane is larger it's a wash.) In testing, there are definitely small detail situations where the 5Ds can lock faster than the R6, even if the detail is not parallel to the R6 AF points and the R6 does lock. If both have good contrast information for their design then it's a wash except with my Tamron 70-200 f/2.8 G2 where the R6 is faster.
There's an interesting inverse to this situation when detail is very low contrast but the area of detail is larger than a single AF point on the 5Ds. (Imagine a large low contrast pattern, something like single color animal fur.) Then the R6 can lock faster. I would guess that mirrorless cameras are always using a relatively wide array of focus points to try and gather enough contrast information to compute a solution, and then perhaps refine the solution with a smaller set as it gets closer. (I comment on this further below.) So in that scenario it's the R6 that has more and better information to start. This is mitigated somewhat by using assist points on the 5Ds.
I also found that with fast primes and long telephotos (i.e. 85mm f/1.4L IS and 100-400L II IS) the 5Ds is less likely to hunt, which the article mentioned. At a certain point of defocus the 5Ds is still able to drive directly to the subject while the R6 has to start hunting to get closer and then drive straight to it. I have one shot of an F-16 Thunderbird during a surprise afterburner flyby where the 5Ds was completely out of focus when I swung around to capture it, and the AF absolutely nailed it before someone's head got in the way, i.e. I had one chance. I've missed a couple shots like that with the R6, and I've gotten into the habit of manually positioning the 100-400 closer to the correct distance. So if a plane just flew by and another one is coming in the distance, I'll swing the focus ring to infinity before activating AF on the R6.
It should also be noted that while the 5Ds AF is good in low light, the R6 kills it here. The R6 amazes me with its ability to focus on stars, for example. The 5Ds can do it, but it needs a very bright one.
Subject Tracking: the 5Ds is actually pretty good here with iTR on. If you put your initial AF point on a face, then it will use face detect to follow it. If it's not a face, then it will just try to follow the shape/color/brightness that was under the selected point. It can drift if there are similar faces or details nearby, but it's definitely usable and 'mirrorless like.'
The R6 is of course better here. Better subject training, faster processor, and more image data. It can hold an eye and is much less likely to drift. Mirrorless passed up the DSLRs that could to this before it could match them on focus tracking (the R for example). One thing that doesn't get enough mention is that 'None' is a valuable setting. If you want to track something that the camera is not trained for, None will try to hold the shape/color/brightness that was under the initial AF point just like iTR DSLRs, only better. Naturally if the camera is trained for a subject it's more likely to hold onto it, but None works pretty good for everything else.
I have found that when photographing pets up close with a fast prime (85mm f/1.4L), I prefer selecting the point. They can move faster than the R6 can compute a solution given that it has to analyze for subject and then acquire focus. If the R6 isn't busy trying to figure out which point to use, then it can keep up. This is a very challenging situation given the focus throw, shallow DoF, and subject speed, and does not reflect on subject tracking under normal conditions which is excellent. Naturally I always use a selected point here on the 5Ds since there is no animal or eye training.
For the heck of it, I'll throw in the EOS 3. While still fast, it's slower than both the 5Ds and R6 at focus acquistion and tracking. Like the 5Ds it's better at acquiring focus when the lens is well out of focus with the center cross point. While it doesn't have modern image based subject recognition, if the subject is well isolated in depth (i.e. a plane in the sky), it can reliably track it across all 45 points.
What does all this mean? As the article points out, DSLRs use purpose built phase detect AF modules. The PDAF module lens assembly and sensor layout are optimized to give actionable contrast information to the computer. This is more than just cross and double cross points, it's also the size of the line pairs. Mirrorless cameras are limited by the size of the AF points, whether they are DPAF (Canon) or PDAF points embedded in the sensor. With the latter if the AF point is too large you have a hole in your image that can't be interpolated away. It took a long time for mirrorless AF to compete with DSLRs on focus tracking, and I would guess that this is the fundamental reason why. I suspect the computer has to look at many points even in 'single point' mode to get the contrast information it needs, and that the computation is more difficult versus the strong signal from a dedicated AF sensor. And with the exception of the R1, the lack of cross points means it can struggle in some situations, namely aligned detail and extreme defocus.
On the other hand, mirrorless has a lot of image detail to use for subject tracking. They're obviously better than most SLRs and DSLRs which can only subject track if the subject is isolated in terms of depth of field, and also better than the DSLRs which have low resolution image sensors for AE meters and for subject tracking. (Though I would be curious to test a 1DX mark III on this, which would be the last best example of a Canon iTR subject tracking DSLR.)
I'd love to get to handle an R1 and put it through the same tests. Faster readout, faster processor, cross sensors, and I imagine it has the faster initial acquisition of the earlier 1DX models. It's a good bet it would clearly beat the 5Ds in all scenarios except maybe the defocus tests where the 5Ds has the double cross sensors.
tl;dr - recent, upper tier DSLRs are very good at focus tracking and remain competitive with mirrorless thanks to dedicated PDAF sensors with large line pairs and cross/double cross points. Canon iTR DSLRs can subject track, but mirrorless is much better at subject tracking thanks to more image data, better training, and faster processing. R1 may very well eliminate the gap in the few places where upper tier DSLRs still have a small focus tracking advantage thanks to the return to cross points.
youtu.be
