Bartlomiej Wilkowski has created a nice tutorial of configuring Java wrapping with VTK.
Java Wrapper Refactoring (Oct 8, 2007)
There were a few problems with the old Java wrappers. One was that, as you said, objects were being deleted before they were supposed to. We hacked in a fix at one point about a year ago which basically made all VTK objects accessed from Java stay around forever, but this was not acceptable either.
The other major concern was that the map from Java objects to VTK objects was in the C++ JNI layer, and while we tried to keep this map synchronized with a mutex, race conditions could still occur because other Java threads could advance while the JNI layer was being called (a thread could access a C++ object just as it is being garbage-collected and deleted). There does not seem to be a way to atomically call a JNI method, or ensure the collector doesn't run while a method is called. This second issue forced us to rethink how the map is done, and the solution was to keep the map in Java instead of C++. But we didn't want this Java map to prohibit objects from being garbage collected. Fortunately, Java has a WeakReference class for just this type of purpose. When accessed, the reference will either be valid or null depending on whether it has been garbage-collected.
Thus, the wrapper code can lookup objects in this map when returning objects from methods, and if it is not there, or null, it creates a new Java object representing that C++ object.
A final issue was that we wanted a way to guarantee all C++ destructors are called before the program exits. The natural place to decrement the reference count of the C++ object is in finalize(), which works when things are garbage-collected, but Java does not guarantee that finalize will ever be called. So the method vtkGlobalJavaHash.DeleteAll() will plow through the remaining VTK objects and call Delete on them.