Animation Database

Multimedia means "numerous mediums" by which information can be stored, transmitted, retrieved and presented. Multimedia databases can support diverse variety of applications in database technology. Some applications are 3D Motion Capture data, data from sensors of gesture sensing devices such as CYBERGLOVE i.e. sensor data, GIS data, stock price quotes etc. The common aspect of such applications is that they involve storage of multiple streams in a unit time and each stream can be considered as an attribute to give multi-attribute characteristics to the database. As the technology is advancing, more and more sophisticated means are developed to generate the repositories which stores multi-dimensional data in huge volumes. In recent years, emerging research in the field of multimedia data is to analyze motion data and retrieve motions efficiently from huge motion database. The important task that lies ahead of researchers is ``Given an query motion, how to search a similar motion in these huge motion databases?''. For instance, 3D motion capture data contains information on the position and orientation of the human body segments by recording the movements of small reflective markers fixed on a human actor while he performs variety of motions (Figure 2). Another interesting application is the ``CyberGlove'', manufactured by Immersion Co-op (Figure 1). As one wears it and performs signs/gestures, each sensor records the movements of fingers and joints of the hand giving multi-attribute characteristic data.

In these motion multimedia applications, more than one value is generated at one instance, unlike time series data sequences where we get only one value at each time. As a result, motion data with many attributes forms a matrix, with rows corresponding to a time axes or frames per second and columns representing each attribute the data.

The main focus of this research is on searching a repository equipped with such multi-attribute motion sequences, for a given motion query by-example using indexing techniques. The main purpose of the indexing is to represent the complex data matrices into representative vectors by preserving the characteristics of the multi-attribute motion sequences and then to prune the majority of the irrelevant motions quickly for a query using index tree. So far no indexing technique can index multi-attribute motion data directly or efficiently.

The matrices of the motion data can be of variable lengths, due to the facts that motions can be carried out with different local speeds and durations and motion sampling rate may also be different. For similar motions, corresponding attributes may have more samples and even difference in values may also be large. Hence, there are no continuous row-to-row correspondences between data of similar motions. These properties of make it difficult to index the multi-dimensional motion data efficiently. Under these difficult circumstances where multi-stream data includes unexpectedly frequent or infrequent co-occurrences, to find a similar match for a given query in such kind of pattern database we need to construct a multi-dimensional indexing structure. On other hand, for handling queries in classical databases there are already many index structures such as R-Trees, SR-Trees, R*- Trees, etc.. These structures are useful for classical databases which have low dimensions but performance of these approaches degrades when applies to the motion data from gesture sensing device or 3D motion data which is high dimensional with multi-attributes features. Hence, our multi-dimensional data in literature is commonly known as data with a ``dimensionality curse''.

NSF CAREER: ANIMATION DATABASE PROJECT

(Funded by National Science Foundation (NSF: http://www.nsf.gov)
through NSF under Grant No. 0237954)

Principal Investigator: Balakrishnan Prabhakaran

OVERVIEW

ANIMATION DATABASES

Animation Toolkit

Reports

Publications