Section: Research Program

Multidimensional indexing techniques

Techniques for indexing multimedia data are needed to preserve the efficiency of search processes as soon as the data to search in becomes large in volume and/or in dimension. These techniques aim at reducing the number of I/Os and CPU cycles needed to perform a search. Multi-dimensional indexing methods either perform exact nearest neighbor (NN) searches or approximate NN-search schemes. Often, approximate techniques are faster as speed is traded off against accuracy.

Traditional multidimensional indexing techniques typically group high dimensional features vectors into cells. At querying time, few such cells are selected for searching, which, in turn, provides performance as each cell contains a limited number of vectors  [84] . Cell construction strategies can be classified in two broad categories: data partitioning indexing methods that divide the data space according to the distribution of data, and space partitioning indexing methods that divide the data space along predefined lines and store each descriptor in the appropriate cell.

Unfortunately, the “curse of dimensionality” problem strongly impacts the performance of many techniques. Some approaches address this problem by simply relying on dimensionality reduction techniques. Other approaches abort the search process early, after having accessed an arbitrary and predetermined number of cells. Some other approaches improve their performance by considering approximations of cells (with respect to their true geometry for example).

Recently, several approaches make use of quantization operations. This, somehow, transforms costly nearest neighbor searches in multidimensional space into efficient uni-dimensional accesses. One seminal approach, the LSH technique  [86] , uses a structured scalar quantizer made of projections on segmented random lines, acting as spatial locality sensitive hash-functions. In this approach, several hash functions are used such that co-located vectors are likely to collide in buckets. Other approaches use unstructured quantization schemes, sometimes together with a vector aggregation mechanism  [96] to boost performance.