They said in the lecture that splitting them more evenly across bounding boxes was a good idea, I'm guessing because we are dealing with tree structures, where a smaller depth and wider tree results in faster computation time across different functions.