approx_screen_percentage = bounding_sphere_radius / (distance_to_object * sin(fov_of_camera))
Isnt this dictated by the longest axis? So imagine you have a column, in fact, a column 50k units tall, standing straight Now, the bounding sphere in this case will be defined by Y. Imagine we are looking at this column with our Camera from the above, completely orthogonal, so you only see the flat end of the column/cylinder. As you go further up by Y-axis the VISIBLE screen size of the object will be decreasing drastically because you only see basically a circle. However, its bounding sphere stays the same and even if we scale it down by some distance-based value it still gonna be min size of longest axis.
So now, if you are far up your actual pixel size is like 20x20 pixels on the screen but because your bound sphere is so big the computed value will be incorrect.