Conclusion - Data Warehouse Systems: Design and Implementation - page 573

data warehouses must be designed in a way similar to traditional data
warehouses. Possible dimensions for image and video data can be the size of
the image or video, the width and height of the frames, the creation date,
and so on. Many of these dimensions also apply to other kinds of multimedia
data.
The main problem in multimedia data warehouses is their high dimen-
sionality. This is due to the fact that multimedia objects like images are
represented in a database by descriptors, which can be of two types: content-
based (or feature) descriptors and description-based (or textual) descriptors.
The former represent the intrinsic content of data (like color, texture, or
shape). The latter represent alphanumeric data like acquisition date, author,
topic, and so on. Most of the content-based descriptors are set oriented rather
than single valued. This would have as a consequence, for example, that we
may need to define each different color as a dimension. Given this high-
dimensional scenario, the main challenge is to be able to perform multimedia
analysis in reasonable execution time.
Image OLAP aims at supporting multidimensional on-line analysis of
image data. An example of the efforts in this field is the work by Jin et
al. [ 97 ], who proposed Visual Cube to perform multidimensional OLAP on
image collections such as web images indexed by search engines, product
images (e.g., from online shops), and photos shared on social networks. Visual
Cube defines two kinds of dimensions: metainformation dimensions such as
date, title, file name, owner, URL, tag, description, and GPS location and
visual dimensions (based on image visual features) such as image size, major
colors, face dimension (indicating the existence of faces), and a color/texture
histogram. To solve the dimensionality problem commented above, the
authors propose two kinds of schemes, namely, a multiple-dimension scheme
(MDS) and a single-dimension scheme (SDS). In an MDS representation, each
possible value of a feature is considered a dimension. For example, Sunny can
be a dimension. Each record corresponding to an image of a sunny day will
contain a '1' in this dimension. On the other hand, in an SDS representation,
the many possible features will be replaced by a dimension denoted Tag .Thus,
an image of a sunny day will contain the value sunny on the Tag dimension.
In addition, a set-valued attribute will contain the identifiers of the images
with that feature, and a single-valued attribute will contain the total number
of such identifiers. The measures in Visual Cube can be a representative
image in a cluster or the number of elements in such a cluster. Clusters
are computed using techniques like the ones studied in Chap. 9 . Records
in a cluster have a combination of descriptors corresponding to the cube
dimensions. In this way, OLAP operations can be performed. For example,
drill-down can be performed by clicking on an image to find others in the
cluster. Open questions are, for example, ecient evaluation of top-k queries
(given a query cell, find the top-k similar cells measured by the similarity