Year: 2016
Type: dissertation
Abstract: This thesis presents the NV-tree (Nearest Vector tree), which addresses the specific problem of efficiently and effectively finding the approximate k- nearest neighbors within large collections of high-dimensional data points. \nThe NV-tree is a very compact index, as only six bytes are kept in the index for each high-dimensional descriptor. It thus scales extremely well when indexing large collections of high-dimensional descriptors. The NV-tree efficiently \nproduces results of good quality, even at such a large scale that the indices can no longer be kept entirely in main memory. We demonstrate this with extensive experiments presenting results from various collection sizes from 36 million up to nearly 30 billion SIFT (Scale Invariant Feature Transform) \ndescriptors. \nWe also study the conditions under which a nearest neighbour search provides meaningful results. Following this analysis we compare the NV-tree to LSH (Locality Sensitive Hashing), the most popular method for ϵ-distance search, showing that the NV-tree outperforms LSH when it comes to the problem of nearest neighbour retrieval. Beyond this analysis we also discuss \nhow the NV-tree index can be used in practise in industrial applications and address two frequently overlooked requirements: dynamicity—the ability to cope with on-line insertions of new high-dimensional items into the \nindexed collection—and durability—the ability to recover from crashes and avoid losing the indexed data if a failure occurs. As far as we know, no other nearest neighbor algorithm published so far is able to cope with all three requirements: scale, dynamicity and durability.
Source: Opin vísindi (Opin vísindi)
Author: Herwig Lejsek
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2338388761