You can do it in RealityCapture, but for such places it could be not so precise. Like, place at least two control points on the images and then you can check the value in pixels and 3D pose of the aligned model.
To reverse the process, I meant to compute it backwards. Like, you know the 2D coordinates, so
(px - (image width / 2)) / scale = m[0]
and so on.
For this we don’t have any documentation, but it is possible to find some way over the internet (I mean the general process).