From the results obtained it seems the usefulness of the global bounding boxes is limited, but this probably has more to do with the nature of the simulation and the behaviour of the objects. In the plant growing simulation, small objects (the vines and leaves) are growing around, and are in close proximity to, large objects. This means that every time the small objects are tested against the large objects they are growing around, there will always be interference between the global bounding boxes. This contributes to the low success rate of the global bounding box scheme. If another type of bounding volume was used, there would still be the same problem.
This highlights the point that the results obtained aren't really a good measure of the collision detection module in general. The results are heavily influenced by the nature of the simulation and the behaviour patterns of the objects. In this simulation, objects are in close proximity to a few other objects, and these neighbouring objects are the only ones that an object needs to be tested against throughout the entire simulation. Also, once a plant part has stopped growing, it no longer needs to be tested against any other part that has stopped growing, or any environment objects. The collision interest matrix is put to extensive use in this simulation to stop checking between the large number of objects where collision is impossible, and also to prevent unnecessary testing of vine segments and leaves that are always in contact.
The poor performance of the global bounding box isn't a crucial problem, as shown by the low voxel per test count --- if the bounding box doesn't cull an unnecessary test, then there is a high probability that the next step of finding the interference submatrices will, and the expensive polygon against polygon testing is still avoided.
Visual inspection of the results shows that the collision detection module appears to be functioning correctly, with no leaves or vine segments growing through any of the environment objects, however more tests need to be carried out with larger databases of environmental objects. This will provide a better test in terms of memory use and requirements. The plants generated may consist of several hundred vine segments and leaves, but they can't quite be regarded as complete objects in their own right in terms of memory usage, because they share the same memory where their vertex and polygon information is kept. Testing large databases would give a better idea in regards to time requirements, as setting up and initializing a large number of objects may take up a significant amount of time.