Research Projects
  In the
Eff2
project we are developing a prototype of an Efficient and Effective image
retrieval system. The retrieval is based on local descriptors, which have been
shown to be a very effective method for image retrieval, e.g. in copyright
detection applications. Our work focuses primarily on 1) The efficiency of the
search process, using advanced techniques to index the massive local descriptor
database; and 2) The user interface, allowing the user control over the
tradeoffs between quality of the result and the speed of the retrieval.
Additionally we have been leading an effort to join together research in image
processing and databases, by founding and chairing the first international
workshop on Computer Vision meets Databases,
CVDB 2004.
The Eff2 project is a cooperation with Laurent Amsaleg of IRISA-CNRS,
Rennes, France and is partially supported by Rannís Technical Research Grant
030290004, EGIDE Jules Verne Travel Grant 4-2003 and several Student Innovation
Fund grants.
Graduate Students: Friðrik Heiðar Ásmundsson, Hafþór Guðnason, Herwig Lejsek,
Vilhjálmur Skúlason
A key consideration for the performance of the database systems, is the speed
of index lookups. This speed depends, among other things on the node size of the
B+-tree indexes employed in all current database servers. Recent developments on
cache-sensitive index structures have the potential to increase the optimal node
size, thereby improving the performance of index lookups. This potential has,
however, not been realized in the research so far. Additionally, these index
structures have not made their way into the main-stream commercial or
open-source database systems yet. The goal of the CSI project is to design and
implement a cache-sensitive index structure within MySQL and evaluate its
performance and its effect on the optimal node size.
The CSI project is a cooperation with Philippe Bonnet of DIKU, Copenhagen,
Denmark and is partially supported by Rannís Research Grant 040044031.
Graduate Student: Árni Már Jónsson
 Semantic
caching is a client caching architecture aimed at supporting selection-based
workloads to servers that support only query based access. Application server
caching is one of the intended application areas. Semantic caching meets the
challenges of today's computing environment by integrating support for query
result caching, reduced network traffic, and application-oriented cache
management policies.
Semantic Caching research has been performed in cooperation with Michael J.
Franklin of U.C. Berkeley, USA, and Divesh Srivastava of AT&T Labs-Research,
New Jersey, USA.
 The Web-Workload Generation project, is aimed at techniques to generate
synthetic web-search workloads that simulate user interests. We have observed
that characteristics of the queries used in web-search (and information
retrieval) studies, significantly affect the performance of various techniques.
In general, randomly generated queries result in performance indicators that are
very different from those resulting from user queries. Extremely long traces,
however, are hard to find. Our project aims at studying the characteristics of a
short trace, and using those characteristics to generate arbitrarily long traces
that make your system behave as is actual users were posing the queries.
 The G3E project focuses on query-by-humming retrieval of music data. The G3E
algorithm is a simple dynamic programming algorithm, that effectively and
efficiently retrieves music that is similar to a hummed song. The project was
partially supported by a Student Innovation Fund grant, and received the
2004 Presidential Student Innovation Award.
Recent Grants
2004-2005 CSI project: Rannís Technical Research fund grant
040044031
2003-2005 Eff2 project: EGIDE Jules Verne travel grant 4-2003
2003-2005 Eff2 project: Rannís Technical Research fund grant
030290003
Student Innovation Fund Grants
2004 Eff2 project: CPU-cache Utilization of an Image Search Algorithm
2004 Eff2 project: Efficient Image Retrieval Using Approximate Cluster Indexes
2003 G3E project: Query-by-Humming (winner of the 2004 Presidential Innovation Award)
2003 Eff2 project: Search in Facial Image Databases
2002 Performance Tradeoffs of Two Database Design Decisions
|
