Database Support for Content-Based Retrieval
For many application domains, similarity search is a quite subjective task for which the user.s preferences should be incorporated. The talk addresses the following aspects:
- (1) Geometric Similarity.
- 3D shape histograms are able to model the similarity of extended spatial objects including protein structures from biomolecular databases or mechanical parts from CAD databases. Quadratic form distance functions help to incorporate the user.s notion of similarity in mind and to cope with small displacements of the shapes.
- (2) Relevance Feedback.
- An important approach to take the user.s needs into account is to iteratively refine the queries by relevance feedback. When considering the cross-correlations of the feature dimensions for the positively marked answers, the query engine has to support quadratic form distance functions with varying correlation matrices.
- (3) Ellipsoid Queries.
- For both concepts, geometric similarity as well as relevance feedback, the quadratic form distance functions yield elliptical query regions. On top of our previous solutions for dynamic ellipsoid queries, we present new approximations that apply to recent vector quantization techniques for indexing high-dimensional feature vectors.
|Published in:||In: Malik J., Kriegel H.-P., Shapiro L., Veltkamp R. (eds.): Content-Based Image and Video Retrieval. Dagstuhl Seminar|
|Forschungsgebiet:||Exploration of Multimedia Databases|