MULTIMEDIA DATABASES

Issues:
Logical Design
Storing and serving multimedia
Retrieval by Content
	  Query Specification
	  Index Structure


Logical Design:
How to fit in relational model?
Use BLOBs or ADTs.  Object-Relational ...
Or use pointers to external files.
Encoding (and compression) are issues.

Physical Design:
Physical layout choices -- within tuple or external
External layout necessary for real-time delivery of large data (video)
Careful placement on multiple disk drives in video-servers.
"Read ahead".
Issue of jitter -- resolve through buffering.
Multiple stream synchronization.


Retrieval by Content:
Use User-Defined Functions in SQL to specify.
These functions may need to refer to other objects.
Provide a friendly interface ...
Multimedia objects often less precisely defined than numbers.
Approximate matches are useful.
  How do you define similarity or measure approximation error?
Two main classes of queries -- (1) similarity  (2) by (computed) attribute
Attributes can be --
	   External  (e.g. date, creator, owner, price)
	   Physical  (e.g. loudness, color histogram, texture)
	   Semantic  (e.g. by visually recognized object in image) - hard to do

Similarity/approximate querying is very hard to specify.
Multiple measures of distance. How to combine?

What is a distance metric?
Notion of transformations.
Data objects as points in a multi-dimensional space.

Attribute Space
  Derive a "feature vector" for the object.
    -K attributes, with a value for each.
  Represent as a point in a K-dimensional attribute space.
  Represent query as a point too.
  Approximation by allowing hyper-sphere around query point.
  Independently bound ranges in each dimension to get query hyper-rectangle.
Feature Vector
  For Image
    -color, possibly by sector
    -brightness
    -texture
  For Document
    -terms that occur in the document

What if Features are Not Known?
  Derive attribute space given only a (dis)similarity function.
  Multi-dimensional scaling.
o	Assign points at random to a space of desired dimensionality.
o	Iterates using steepest descent to minimize "stress" function.

Index structures to support efficient access -- filter, not solve.
      Thumbnails -- user picks
      More expensive analyses -- software picks


MULTI-DIMENSIONAL INDEX STRUCTURES:

Applicable even in non-spatial applications.

R-trees.
   First decent multi-dimensional structure.
   Many embellishments since.
   Understand how it works in terms of GiST.

The curse of high-dimensionality.
  Most multi-dimensional index structures do not scale to high dimensions.
  Basic intuitions break down
-	More corners than points in the data set.
-	Low selectivity in any one dimension.
-	"Center point" included in most queries.

Dimensionality Reduction
  Data often is not "truly" high dimensional.
  Variety of transforms can be used.
-	Fourier transform
-	Wavelet transform
-	Singular Value Decomposition
  And coefficients with small (absolute) values can be dropped.