"By using a vocabulary adapted to the likely distribution of data, we use a much smaller tree, resulting in better resolution while maintaining a compact representatin."

When the authors pick a vocabulary to adapt to the likely distribution of the data, are they doing that for every on-the-fly inserted object? I was thinking that each time a new object is inserted into the tree on-the-fly, they can check its distribution first to find the tree that has the vocabulary most like it.