Fast Filtering for Nearest Neighbor Search by Sketch Enumeration Without Using Matching

A sketch is a lossy compression of high-dimensional data into compact bit strings such as locality sensitive hash. In general, k nearest neighbor search using sketch consists of the following two stages. The first stage narrows down the top K candidates, for some K = k, using a priority measure of sketch as a filter. The second stage selects the k nearest objects from K candidates. In this paper, we discuss the search algorithms using fast filtering by sketch enumeration without using matching. Surprisingly, the search performance is rather improved by the proposed method when narrow sketches with smaller number of bits such as 16-bits than the conventional ones are used. Furthermore, we compare the search efficiency by sketches of various widths for several databases, which have different numbers of objects and dimensionalities. Then, we can observe that wider sketches are appropriate for larger databases, while narrower sketches are appropriate for higher dimension.