Approximate box decomposition trees--近似盒分解树

Approximate box decomposition trees

Arya et al. [4] have presented an optimal algorithm for approximate nearest neighbor search. They use a balanced box decomposition tree (bd-tree) as their primary data structure. This tree combines two important properties of geometric data structures: First, as in the d-tree case, the set of points is exponentially reduced. Second, the aspect ratio of the tree edges are bounded by a constant. Not even the optimized d-tree is able to make this assurance, but quadtrees show this characteristic [4]. The actual box decomposition search tree is composed of splits and shrinks. Fig. (c) shows the general structure and Fig. presents two slices within this search tree.

Figure: Left: The -approximate nearest neighbor. The solid circle denotes the environment of . The search algorithm need not analyze the gray cell, since satisfies the approximation criterion. Middle and right: (a) Given point set. (b) decomposition into buckets. (c) Tree layout. Fig. adapted from [ 3, 4].

Figure: A bd-tree of scanned 3D data ( ) of Fig. (top) first/black 3D scan. Two -projections of slices at depths cm (a) and cm are given (b).

(a) (b)

The search procedure of bd-trees is similar to the one of d-trees. The approximate search is discontinued (cf. Fig. ) if the distance to the unanalyzed leaves is larger than

来源： http://kos.informatik.uni-osnabrueck.de/download/icar2005_2/node17.html