Or maybe it is because it isn't actually mainly opencv code and the top comment didn't actually look at it but nobody checked and just upvoted.

One thing that ARKit is able to do is utilize information about the device motion from the accelerometer and gyroscope. Also, I believe it is a slightly different technique than what is described in this post. I think they do more of a 3D reconstruction (i.e. structure from motion), and build up a sparse model of the environment, kind of like SLAM. That can perform better because it uses the knowledge that the environment is relatively constant between frames. This post, and most examples you see, essentially start over from scratch each frame, throwing away all the information from the last tracking frame.

Of course, many of the algorithms you need to do that are part of OpenCV, so maybe it is still 95% OpenCV!

Probably a combination of not using something as simple as RANSAC, and passing the output through a Kalman filter to smooth it out. (so: sufficiently advanced glue until proven otherwise)

That’s certainly true.

The next thing I will do if I keep working on it will be to implement a Kalman filter to reduce shakiness. It could visually improve a lot the results since coherence of the projection between frames would make it much more attractive.

And I think this is easier to achieve with a Kalman filter rather than using a more precise estimation method.

Yeah, that alone might do it. Only one way to find out :-)

>5% novelty