Poster Submissions


History and Foresight for Distance-Based Relevance Feedback in Multimedia Databases

Marc Wichterich, Christian Beecks and Thomas Seidl
Data management and data exploration group RWTH Aachen University, Germany

Users of multimedia databases often face the difficulty of expressing similarity queries in a fashion that can be directly utilized by information retrieval systems. While users discern relevant from non-relevant objects based on high-level concepts, retrieval systems make this decision based on lower-level features. A promising approach to overcome the semantic gap between subjective user preferences on the one hand and feature-based, quantitative measures on the other hand are distance-based relevance feedback systems. In each iteration of the feedback loop, database objects marked as relevant by the user are employed to derive a new distance function that attempts to increasingly reflect the user's notion of similarity. Current distance-based relevance feedback systems derive such a distance function from the relevance information of the last iteration only. In contrast, we propose to utilize a weighted history of all objects that the user selected as relevant throughout the entire relevance feedback procedure in order to more accurately reflect the userís preferences. We show how the history information can be efficiently introduced to a popular relevance feedback framework based on quadratic forms. Using the same mathematical framework, we also enable a seamless transition between a phase of quickly approaching the region of relevant objects in a database (exploration with foresight) and a phase of adapting to the local preferences in that region.

Hosted by the Knowledge Media Institute