Improving Affinity Matrices by Modified Mutual kNN-Graphs

The recent progress in describing affinities between images or objects by means of shape, appearance or texture allows the exploitation of inherently emerging redundancies for improvement of retrieval tasks. We propose a two-way normalization and analysis scheme which aims on (a) modeling object interdependence by neighborhood incorporation and (b) retrieval improvement by subsequent analysis from a modified mutual k nearest neighbor graph. We provide a general and flexible approach which may be either applied for improving retrieval quality or as base for semi-supervised classification, clustering or dimensionality reduction methods. The presented experiments demonstrate that our approach yields to significant improvements on a broad variety of data sets, including the highest ever reported bullseye score of 93.40% on the MPEG-7 database.