Selecting vantage objects for similarity indexing

Abstract

To make similarity searching in multimedia databases practical, indexing has become a necessity. Vantage indexing is an indexing technique which maps a dissimilarity space onto a vector space such that each object is represented by a vector of dissimilarities to a small set of m reference objects, the vantage objects. Querying takes place within this vector space, reducing the number of distance calculations to m. The retrieval performance of a system based on this technique can be improved significantly through a proper choice of vantage objects. We propose a new technique for selecting vantage objects and present experimental results based on data sets of different modality.

Efficient Metric Indexing for Similarity Search and Similarity Joins

“…Venkateswaran et al [35] choose pivots that maximize pruning for a sample of queries. Leuken and Veltkamp [36] select pivots with the minimum correlation to ensure that objects are evenly distributed in the mapped vector space. Recently, PCA [37] has been developed for pivot selection.…”

“…9 shows the results obtained using real datasets. The first observation is that HFI performs better than the existing pivot selection algorithms considered, i.e., HF [6], Spacing [36], and PCA [37]. The reason is that the search performance is highly related with the precision as defined in Definition 1, and HFI tries to maximize precision.…”

Some vantage indexing approaches for rhythmic and melodic search

“…To construct still better vantage indices without making the construction process prohibitively time-consuming, one can use different properties of vantage objects, and one can also try to measure the quality of a whole vantage index instead of just optimizing certain properties of additional vantage objects. One such method, which inspired the new approaches described in the next section, is to pick vantage objects based on a spacing criterion and a correlation criterion (Leuken, Veltkamp, & Typke, 2006).…”

“…Algorithm 1 is a slightly modified version of the algorithm from (Leuken et al, 2006) for selecting a good set of vantage objects from the database to be indexed. The most computationally expensive step in each iteration is to calculate the distances between a new vantage object candidate c and all database objects.…”

“…If the thresholds were picked loosely enough, it will not only return a result but also make the requirements more stringent once a sufficient set of vantage objects has been found, and continue looking for a better set. The automatic adjustment of the thresholds is one of the modifications in comparison to the algorithm from (Leuken et al, 2006).…”