of course, aligning the virtual camera of the 3D scene with the real camera’s capture of the actual scene was never going to be straightforward. easy to get to a proof of concept. hard to actually register the two. i ended up rendering a cuboid grid on the seat positions in the 3D scene, drawing by hand (well, mouse) what looked about right on a video still, and trying to match the two sets of lines by nudging coordinates and fields-of-view with some debug-mode hotkeys i hacked in.

in hindsight, i would have stuck motion capture markers on the cameras. so it goes.