Performance evaluation of visual features for
structuring of video sequences and content-based image
Download Bibtex entry
||Technische Universität Berlin
thesis considers the performance evaluation of 44 visual features, 27
distance measures and two methods for the combination of features and
distances for the cluster analysis of extracted keyframes from videos
as well as for content-based image retrieval. For the evaluation of
the cluster analysis, a metric Variation of Information (VI) has been
considered. Content-based image retrieval has been evaluated according
to measures Average Normalized Modified Retrieval Rank (ANMRR) and
Mean Average Precision (MAP). Improvements in clustering results as
well as in retrieval results can be achieved by fusing features. The
results can be improved by the combination of visual features. For
that a supervised sequential forward selection of features is used.
Another way to improve the results is the weighted combination of
distances of different visual features. This thesis deals with the
approaches for the automatic determination of these weights.
Therefore, several objective measures from graph theory have been
examined to what extent they are suitable for an unsupervised
evaluation of clustering results.