We have found that most important for the retrieval quality is to have a large vocabulary (large number of leaf nodes), and not give overly strong weights to the inner nodes of the vocabulary tree. In principle, the vocabulary size must eventually grow too large, so that the variability and noise in the descriptor vectors frequently move the descriptor vectors between different quantization cells. The trade-off here is of course distinctiveness (requiring small quantization cells and a deep vocabulary tree) versus repeatability (requiring large quantization cells). However, a benefit of hierarchical scoring is that the risk of over doingthe size of the vocabulary is lessened. Moreover, we have found that for a large range of vocabulary sizes (up to somewhere between 1 and 16 million leaf nodes), the retrieval performance increases with the number of leaf nodes. This is probably also the explanation to why it is better to assign entropy directly relative to the root node. The leaf nodes are simply much more powerful than the inner nodes.
đang được dịch, vui lòng đợi..
