|
I don't quite understand the purpose of having multiple semi randomized kd-trees instead of just having one. vlfeat's documentation says that it is for the purpose of improving the searches at higher dimensions. But are the payoffs worth it? It seems like tree queries end up being slower as a result. |
|
I would imagine the benefit is similar to that of having an ensemble of decision trees rather than just one: The representation is more compact. Partitioning a space using a single tree can cause pathological structures, whereas overlaying several trees can model the space more naturally. Let me know if you want me to dig up a reference or a visualization of this for you. |
