Visual feature space analysis SPP 1335
Many important data retrieval and analysis tasks such as similarity search, clustering, and classification, require feature vector (descriptor) representations to calculate distances between instances of complex data types. However, it is usually a priori not clear what the best feature descriptor to solve a given application problem is. In most cases, a wealth of options for generating descriptors exists, including choice of feature type, and setting of preprocessing, normalization, and level of detail parameters. Above all, the feature vector descriptors to be employed need to fit the given application domain, data set characteristics, and user task. Only feature vectors appropriately configured to the overall application, user, and data environment will yield satisfactory analysis results. Offline benchmarking on predefined reference data sets is of limited use, as it is data- and application-dependent, and expensive in terms of supervised information required.
Located in the fields of data analysis and visualization, this project aims at generically improving the performance of feature-based data analysis applications, by means of novel tools for interactive and user-centered visual feature space analysis. Effective Visual Analytics methodology will be developed to support visual-interactive analysis and configuration of candidate feature vector descriptors, to best fit the given data, application, and user task at hand. Integrated with respective feature-based applications, automatic feature analysis and configuration methods will be combined with appropriate visualization and interaction facilities. Automatic analysis will screen the space of possible feature descriptors, identifying promising candidate features. Visualization of the automatically obtained analysis results, in combination with appropriate user interaction and feedback, will allow converging to a set of feature descriptors best supporting the user task at hand. The Visual Analytics methodology to be developed will be embedded into real-world, relevant types of applications and data.
Presentation Slides from Dagstuhl Seminar on Scalable Visual Analytics, Nov. 2010:
- JP Schreck (Visual Analytics)
|Period:||18.08.2008 – 17.08.2011|