Instead of using the built-in camera fitting function,
which wants to rotate the camera, it now calculates the
local bounding box of the object for each camera angle.