Nearest neighbor searching in high dimensions using multiple KD-trees

  • Authors:

  • Shwu-Huey Yen;Chao-Yu Shih;Hsiao-Wei Chang;Tai-Kuang Li

  • Affiliations:

  • Department of Computer Science and Information Engineering, Tamkang University, Tamsui, Taipei County, Taiwan;Department of Computer Science and Information Engineering, Tamkang University, Tamsui, Taipei County, Taiwan;Department of Computer Science and Information Engineering, Tamkang University, Tamsui, Taipei County, Taiwan and Department of Computer Science and Information Engineering, China University of Sc ...;Department of Computer Science and Information Engineering, Tamkang University, Tamsui, Taipei County, Taiwan

  • Venue:
  • ISCGAV'10 Proceedings of the 10th WSEAS international conference on Signal processing, computational geometry and artificial vision
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Feature matching plays a key role in many image processing applications. The matching efficacy is crucially determined by the description of features. To be robust and distinctive, feature vectors usually have high dimensions. Thus accurately finding the nearest neighbor of a high dimensional query point in the target image becomes essential. In this paper, we propose a multiple kd-trees method to locate the nearest neighbor for high dimension feature points. First, we project feature points to three hyper-planes corresponding coordinate axes that are with the first three greatest variances. Secondly, for points projected on each splitting hyper-plane, two kd-trees are built that one is the conventional kd-tree and the other has a first split on the hyper-plane with the second largest variance. Thus in total six kd-trees are built to compensate the deficiency of projection may have. Although our method requires a longer time in tree construction, it is still quite efficient (not more than 0.62 seconds for 1000 data of dimension 50). The experiment showed that our method improves the precision of the nearest neighbor searching problem. When the dimension of data is 64 or 128, the average improvement on precision can reach 28% (dimension fixed) and 53% (number of backtracking fixed).