It depends: Culling on the clipping planes (not just the far one) happens per object (based on its bounding box), per octree-node and finally per polygon. The most efficient of these are the octree-nodes, because you can cull away very large parts at once. A 50.000 polys scene with 5000 visible polygons can't perform as good as a scene with only 5000, but the overhead (with octrees) should be neglectable. When splitting the scene into single objects, the overhead is larger because there is no hierarchy information like there is in the octree. Culling individual polygons is the killer(tm) or course.
so this may well worth a try
![Wink ;)](https://www.jpct.net/forum2/Smileys/default/wink.gif)
merge the whole scene into a single object, use octrees on it, set clipping plane to a reasonable value, place LoD planes just at clipping plane's border.
the missing part is, i dont want to/cant merge the whole scene into a single object because of our infamous climbing problem. i do use collison meshes for buildings, trees etc. that's i need them as seperate objects. maybe implementing an inter-object octree system and setting far ones invisible may help, indeed setting them invisible depending on a simple distance check may also help..