Alright I took another look at the Adaptive Binary Tree because when I tested the octree implementation on a level over 250.000 polygons and about 30 different materials the octrees performance degraded. When I first implemented the ABT I also tested on this level but I did not fully implement materials because I did not expect that it would affect performance a lot. When I was implementing materials into the octree implementation though, the performance dropped very fast because sets of polygons with the same material are fragmented throughout the level. Due to the nature of the octree lots of small sets are created.
The ABT however is able to adapt much better to this. So I took another shot at it. This time I also created a much better compiler! I attached a screenshot of the ABT in maya. I must note that creating the ABT takes about 30(!) seconds. There is also the screenshot of the temple so you can compare the result of the octree. (Compiling the ABT for the temple scene takes less then a second.)