and a Boolean value indicating whether or not the image is a landscape. Such a set of attributes is referred to as a As in conventional database management systems (DBMSs), to allow users to efficiently access and retrieve data objects, a feature vector. When represented in this manner, each immultimedia database management system (MMDBMS) must age becomes a point in a k-dimensional space, where k is employ an effective access method such as indexing and hashthe number of features in each vector [Fal96]. Figure 1.1 ing. This paper provides a survey of tree-based multidimenillustrates this using a three-dimensional feature vector. sional indexing techniques for MMDBMSs that maintain image The vector contains the values 3.5 in the X-dimension, 0 data represented as feature vectors. These techniques support in the Y-dimension, and 8 in the Z-dimension. such data while maintaining desirable characteristics of a B-Such a feature vector can be created for any image by tree, an index structure most commonly used in traditional applying some feature extraction algorithms to it. When DBMSs. In this survey, we provide descriptions of each tree the same group of general algorithms are applied to the as well as give examples of the different data organization set of images in an MMDBMS, the images will then be schemes. We also describe the advantages and disadvantages represented as feature vectors with the same dimensions. of using each technique. In addition, we provide classifications This means that they can be treated as different points in of the trees using several different properties. These classificathe same k-dimensional space.
tions should assist researchers in identifying the strengths and
When developing indexing methods for image data, weaknesses of any new indexing technique they develop as well then, researchers often operate under the assumption that as help users determine the most appropriate data structure the images are represented as feature vectors in the same for their applications.