Ytukyg

The setup is this: we're trying to show an AR object in an open space standing next to a marker. The purpose of the marker isn't to position the object 100% on/at the marker but to give a "best guess" that the person is looking in the right direction.

Imagine the viewer looking at a scene where there's a traffic sign, and they are looking at it from e.g. 15 yards away, and while the person is looking at the traffic sign, we want to draw an AR character approximately next to the sign. Just instead of the traffic sign (but approximately of the same dimensions) we can use an arbitrary image.

While viewing this scene through the AR camera from this distance, any "natural looking" marker is rather small.

Is there a way / technique / best practice of picking / adjusting / optimising a marker so that it's detected while occupying a relatively small pixel area from the camera?

And an additional question: Naively, without much knowledge in this field, I'd expect that the best markers would be a high contrast text, or a non-symmetrical arrangement of geometrical shapes (so, literally like traffic signs), but apparently more realistic images with shades in them are preferred? Why is that so?

Answer to your first question is really not easy. Because it is not just detecting a marker or an augmented image but it is tracking these images. So you can gather position and rotation information of the marker as well. Any change in the rotation and the position of the marker can be detected. This is essentially done via extracting feature points and tracking position of these points relative to different pose of the device. My advice to you could be trying HelloAR example and trying to visualize feature points on your marker from 15 yard. This will give you the rough idea why you can not optimize such distances. ARCore is using visual odometry so these feature points play a vital role in tracking. This is probably the reason to your additional question because we do not only need to detect the markers but we also have to track them. Therefore, more optimal points means more feature points and this makes tracking easier. One bad work around might be if you know the distance of the traffic sign and start your AR app aligned with the sign you can create arbitrary anchors at any points using `Session.createAnchor. However, as i said this will not be a good solution. I can try to help if you provide more information about the use case maybe somehow you can make arbitrary anchors work. Good luck!