Using human partitionings of image sets to learn a similarity-baseddistance measure for the organization of image databases

Bibtex entry :

@inproceedings { VG:Squ1998c,
    author = { David McG. Squire },
    title = { Using human partitionings of image sets to learn a similarity-baseddistance measure for the organization of image databases },
    booktitle = { Multimedia Storage and Archiving Systems III (VV02) },
    pages = { 80--88 },
    year = { 1998 },
    volume = { 3527 },
    series = { SPIE Proceedings },
    address = { Boston, Massachusetts, USA },
    month = { November },
    keywords = { image databases, database organization, similarity, learning, agreement },
    note = { (SPIE Symposium on Voice, Video and Data Communications) },
    abstract = { In this paper our goal is to employ human judgments of image similarityto improve the organization of an image database for content-basedretrieval. We first derive a statistic, $\kappa_B$ for measuringthe agreement between two partitionings of an image set into unlabeledsubsets. This measure can be used both to measure the degree of agreementbetween pairs of human subjects, and also between human and machinepartitionings of an image set. It also allows a direct comparisonof database organizations, as opposed to the indirect measure availablevia precision and recall measurements. This provides a rigorous meansof selecting between competing image database organization systems,and assessing how close the performance of such systems is to thatwhich might be expected from a database organization done by hand.We then use the results of experiments in which human subjects areasked to partition a set of images into unlabeled subsets to definea similarity measure for pairs of images based on the frequency withwhich they were judged to be similar. We show that, when this measureis used to partition an image set using a clustering technique, theresultant clustering agrees better with those produced by human subjectsthan any of the feature space-based techniques investigated. Finally,we investigate the use of machine learning techniques to discovera mapping from a numerical feature space to this perceptual similarityspace. Such a mapping would allow the ground truth knowledge abstractedfrom the human judgments to be generalized to unseen images. We showthat a learning technique based on an extension of a Kohonen networkallows a similarity space to be learnt which results in partitioningsin excellent agreement with those produced by human subjects. },
    owner = { steph },
    timestamp = { 2008.05.04 },
    vgclass = { refpap },
    vgproject = { viper },
}