1998
@inproceedings{Squ1998c,
vgclass = {fullconf},
vgproject = {cbir,viper},
author = {David McG. Squire},
title = {Using human partitionings of image sets to learn a
similarity-based distance measure for the organization of image
databases},
editor = {C.-C. Jay Kuo and Shih-Fu Chang and Sethuraman
Panchanathan},
booktitle = {Multimedia Storage and Archiving Systems III (VV02)},
address = {Boston, Massachusetts, USA},
volume = {3527},
series = {SPIE Proceedings},
pages = {80--88},
month = {November},
year = {1998},
note = {(SPIE Symposium on Voice, Video and Data Communications)},
abstract = {In this paper our goal is to employ human judgments of
image similarity to improve the organization of an image database for
content-based retrieval. We first derive a statistic, $\kappa_B$ for
measuring the agreement between two partitionings of an image set into
unlabeled subsets. This measure can be used both to measure the degree
of agreement between pairs of human subjects, and also between human
and machine partitionings of an image set. It also allows a direct
comparison of database organizations, as opposed to the indirect
measure available via precision and recall measurements. This provides
a rigorous means of selecting between competing image database
organization systems, and assessing how close the performance of such
systems is to that which might be expected from a database organization
done by hand.
We then use the results of experiments in which human subjects are
asked to partition a set of images into unlabeled subsets to define a
similarity measure for pairs of images based on the frequency with
which they were judged to be similar. We show that, when this measure
is used to partition an image set using a clustering technique, the
resultant clustering agrees better with those produced by human
subjects than any of the feature space-based techniques investigated.
Finally, we investigate the use of machine learning techniques to
discover a mapping from a numerical feature space to this perceptual
similarity space. Such a mapping would allow the ground truth knowledge
abstracted from the human judgments to be generalized to unseen images.
We show that a learning technique based on an extension of a Kohonen
network allows a similarity space to be learnt which results in
partitionings in excellent agreement with those produced by human
subjects.},
}