High-Dimensional Indexing for Multimedia Features

Efficient content-based similarity search in large multimedia databases requires efficient query processing algorithms for many practical applications. Especially in high-dimensional spaces, the huge number of features is a challenge to existing indexing structures. Due to increasing overlap with growing dimensionality, they eventually fail to deliver runtime improvements. In this work, we propose an overlap-free approach to indexing to overcome these problems and allow efficient query processing even on high-dimensional feature vectors. Our method is inspired by separator splits e.g. in B-trees for one-dimensional data or for sequence data. We transform feature vectors such that overlap-free splits are ensured. Our algorithm then queries the database with substantially reduced number of disk accesses, while ensuring correctness and completeness of the result. Our experiments on several real world multimedia databases demonstrate that we build compact and overlap-free directory information in our index that avoids large percentages of disk accesses, thus outperforming existing multidimensional indexing structures.