Thom Hogan wrote:

Laqup wrote:

the golden middle has been used ("3"). To be honest, at least according to the decsription and basic algorithm design principles it should not matter, as there is nothing moving in between subject & camera and actual subject detection has never been lost.

I'm going to have to go to the replay booth on this one ;~). One thing I believe I noticed in your examples—hard to tell for sure because of the way you presented them—is that the camera is indeed moving. I've written this before—and to some degree it's one of the things that is a bit different between brands—but handling discipline informs AF on the Nikons, for sure.

Maybe I phrased that wrong, with moving I actually meant "passing". The camera has definitively moved and the subject as well. I mostly was sitting / kneeing somewhat stationary, but there is always some movement. This is what I hoped would be tackeled by AF-C and subject detection (aka constant refocusing or at least focusing on shutter half press). With Canon it works that way.

But: As far as I understand the option "A3", blocked shot is not related to camera or subject movement but rather to objects that PASS in the line of sight between camera and subject. In my examples there was nothign passing in between but I always had a "unobstructed" line of sight. Please correct me if I understand the block shot option wrong.
My interpretation is, that in case of "5" (aka delayed), a passing object would not lead to an immediate switch of the subject, the camera will rather try to "predict" / "estimate" / "guess" where the previous target was, where a setting of "1" (quick) would immediately lead to a switch of the subject from original to the one passing in between.
I am an algorithm developer in the field of computer vision and this is a typical parameter for subject tracking, where the implemented filter ("predict / update" cycle) in case of occlusion / missing measurement would stay longer in the "prediction" mode w/o increasing the underlying variances that allow measurement association or alternatively w/o deleting the track.

In my world to allow a focus switch from the eye to the hand (if the hand is being interpreted as a passing object) would first have to be triggered by a loss of the actual subject/eye detection, which at least was not indicated in the GUI.
If the camera is performing a full area auto AF point selection parallel to subject detection and then even grants priority to this "hidden auto AF point selection" even if the original subject is still present and has a valid measurement this is a rather weird algorithm implementation. Maybe I should apply at Nikon to get things finally right then.

Here a typical example where I assumed "blocked shot" would be relevant:

In image #14 the person is detected and properly tracked, in image #72 the object is occluded by a "pseudo passing object". If blocked shot is set to quick, the track would be deleted very fast and either the lamp post or the second person would be selected as new subject. With blocked shot set to "delayed" the algo would predict were the person went and "wait" a few frames before changing focus distance / deleting the object or selecting a new subject.
I am not sure if the subject / eye detection is actually being tracked at all though, to me it looks more  like a "single cycle" event, maybe with a very short prediction window. 3D tracking is a different beast though.