Thanks Ian. Your explanation is insightful but my question was really about the AF mode used for this observation: " AF accuracy and consistency in view-finder mode is about the same as I see from my 5Ds and 1Dx II cameras and the AF speed seems comparable to what I get from the bigger cameras and lenses". It sounds like it might have been AI Servo?

As for the variation between shots in AF, I think this relates (but not sure why) to the sensor and/or algorithm operating when not in live view. When AF is done in live view, there is no microadjustment applicable, of course, because the image is being focused on the main imaging sensor, not a secondary sensor. In my limited experience, live view AF is invariably spot on although I haven't done a lot of testing. So I'm not sure it is simply depth of field at play here.

No Glenn, I am talking about the experience when using my AFMA rig and one shot mode. I manually defocus the lens then press the shutter button (or remote release in the case of 7DII, 5Ds and 1Dx II). The speed of focus is difficult to judge, what I am saying is that it seems to take about the same amount of time for the combinations I have tested. So for example, the 90D + 400 mm f/4 DO II seems to snap onto focus in about the same time it takes my 1Dx II + 600 mm f/4 III to snap onto the focus. One is not obviously faster than the other. The variation in the accuracy of focus (front to back) as a percentage of the target distance is also similar keeping in mind that I test the 1Dx II + 600 mm f/4 III at a range of about 25 m. And yes, Live View is much more consistent, that is not an issue.
Cheers, Ian

Thanks Ian for the useful observations. I expect the comments going to the comparison between different models will be useful for 7DII users

I spent some time tracking gulls on Lake Burley Griffin several days ago, with similarly mixed results, using the 9 point autofocus system and the highest continuous shutter repeat. Some sequences were quite successful in tracking the bird and maintaining good focus, others not so even when the position of the bird in the frame is under one or more of the 9 points. Interestingly I had better luck with birds flying towards the camera than with those panning horizontally. However, I am unable to draw useful conclusions because I have no way of knowing how much this was due to the camera and how much to me, noting that I am not particularly practised at BIF and its a skill that needs to be cultivated. I want to get down to try again with some different camera settings but with a string of 40 degree C days predicted ahead, plus the looming childcare duties of the school hols I am not sure when the opportunity will arise.

In a sense, this is academic; its interesting checking out the limitations of the camera in this particular type of photography but my frame of reference for what is 'good' in this regard is limited, and besides the camera is a keeper overall.

On the autofocus subject Roger Cicala's series of articles on Lensrentals.com comparing (fairly informally) the performance of phase detection AF systems with contrast detection AF on a range of cameras and lenses, is interesting reading. While these articles are now fairly old, he uses resolution graphs to show the differences between the contrast detection and phase detection with particular Canon lenses and bodies. Using contrast detection (live view), and also using manual focus with magnified rear screen images, he was able to achieve good consistency of focus over several shots. There are still differences in resolution from shot to shot but these are comparatively small. In contrast, phase detection showed rather more substantial variation between shots, and while most shots were still within acceptable limits of focus there were still a reasonable proportion where the focus would be deemed unacceptable. Autofocus microadjustment was able to correct lenses which were obviously so off that no shots were acceptable, but did not correct the shot to shot variance. Moreover there was substantial variation between different bodies and lens combinations, generally with improvements with newer technology.

With wider reading, it seems that there are potentially a range of reasons for the variation of phase detection from shot to shot, some of which I don't really understand. Clearly the physical calibration of the system has to be very exact, and can vary with (for example) temperature. In addition, I understand that an adjustment needs to be made to take account of the differences in how light focuses on the main sensor compared to the narrow-apertured microlenses on the autofocus sensor, leading to an estimated 'best focus correction value' which is hard coded into firmware in the lens and must be applied by the camera's computer. Things like aperture and available light also affect the accuracy of phase detection. Canon (and I presume other manufacturers) also use predictive algorithms to have the camera anticipate movement of the subject during high speed shooting, which are likely to be 'deceived' by some situations, a situation which Canon recognises by having different settings for prediction based on different types of subject movement. Canon highlights in its web articles the fact that a subject moving at speed will move substantially (sometimes metres) in the short time between pressing the shutter button and the shutter activating (shutter lag) as well as between shots, and also the camera cannot correct focus when the mirror is up. So when you think about how much time the camera has to focus in between the mirror flipping up and down up 10 times a second, and the complexity of the decisions to be made to predict subject movement, I'm not surprised the camera sometimes misses focus during high speed shooting.

Returning to the 90D, I continue to be impressed overall by the camera. Now that DxO Photolab has been updated for the 90D, I am able to use my preferred processing settings in that software with very pleasing results, although I haven't done any really high ISO work. A solid upgrade on the 80D, and a good all round performer which balances well with the 100-400 II lens.

Simon

Over the last few days I have revisited the AFMA issue in the hope that I might find more quantitative answers to some of the questions raised by Glenn et al. I did a quick and dirty AFMA with the 300 mm f/2.8 II and 400 mm f/4 DO II a couple of weeks ago so that Jill and I could use the new camera with some confidence that the birds would be in focus. Over the last few days I have used the 90D + 300 mm f/2.8 II to investigate the AF performance more carefully using my AFMA test rig described earlier in this thread. In particular, I wanted to measure the consistency of focus using One Shot AF and AI Servo AF. To do this I measured the amount of defocus from the ideal position using a series of 51 shots. The AFMA used was the value determined by my earlier quick check using 8 shots; this value was +2 AFMA units. The results from the long series of shots confirm this was the correct AFMA to apply and also showed the frequency distribution of the defocus. It also showed that there was little difference between the consistency of focus achieved with One Shot AF and AI Servo AF.

During the course of this tedious investigation I became very familiar with the appearance of the test target and came to a conclusion about what was acceptable sharpness, at least, for my eyes. The resolution test patterns on my target are very sensitive to defocus and using these I decided that provided the defocus was no more than about +/- 4 cm, the test target was acceptably sharp. You will immediately perceive that this is way less DoF than what would be suggested by a DoF calculation. This is because DoF calculators estimate the DoF for prints viewed from a certain distance which in no way corresponds to viewing a computer monitor at normal working distance. After having decided that a defocus of +/- 4 cm was acceptable, I was then able to calculate the percentage of sharp images using the full data set from 102 images. The result was that 83% of shots were acceptably sharp, not a bad result I think. I have attached the defocus frequency distribution below so that you can see for yourself what the results of this experiment look like. I am contemplating whether to repeat the experiment for the 90D + 400 mm f/4 DO, it may tell us something about the AF consistency for an f/4 system, I anticipate there will be more shot-to-shot variation.

Thanks for that detailed investigation and having the patience to do it. The message I take away from this is to take several bursts of shots for subjects in case the AF is out for one of the bursts (especially if using single shot AF). An advantage of AI Servo will be that a single burst should be OK since the camera should try to refocus after every shot. Based on your results, one in roughly six shots should be out of focus, or at least not acceptably sharp. Very useful info and thanks again.

Thanks Ian, interesting results. I would imagine that this is quite consistent with cameras of this type and in this price bracket. Unfortunately you have traded in your 7D II or you could have compared to see if it is any better. Another interesting comparison (not that I am suggesting you actually do it!) would be to do the tests with Live View - I expect that you would see a similar spread but perhaps with the outlier values closer to zero (ie more consistency).

Simon