Computer Laboratory

Technical reports

Evaluating similarity-based visualisations as interfaces for image browsing

Kerry Rodden

September 2002, 248 pages

This technical report is based on a dissertation submitted 11 October 2001 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Newnham College.

Some figures in this document are best viewed in colour. If you received a black-and-white copy, please consult the online version if necessary.

Abstract

Large collections of digital images are becoming more and more common, and the users of these collections need computer-based systems to help them find the images they require. Digital images are easy to shrink to thumbnail size, allowing a large number of them to be presented to the user simultaneously. Generally, current image browsing interfaces display thumbnails in a two-dimensional grid, in some default order, and there has been little exploration of possible alternatives to this model.

With textual document collections, information visualisation techniques have been used to produce representations where the documents appear to be clustered according to their mutual similarity, which is based on the words they have in common. The same techniques can be applied to images, to arrange a set of thumbnails according to a defined measure of similarity. In many collections, the images are manually annotated with descriptive text, allowing their similarity to be measured in an analogous way to textual documents. Alternatively, research in content-based image retrieval has made it possible to measure similarity based on low-level visual features, such as colour.

The primary goal of this research was to investigate the usefulness of such similarity-based visualisations as interfaces for image browsing. We concentrated on visual similarity, because it is applicable to any image collection, regardless of the availability of annotations. Initially, we used conventional information retrieval evaluation methods to compare the relative performance of a number of different visual similarity measures, both for retrieval and for creating visualisations.

Thereafter, our approach to evaluation was influenced more by human-computer interaction: we carried out a series of user experiments where arrangements based on visual similarity were compared to random arrangements, for different image browsing tasks. These included finding a given target image, finding a group of images matching a generic requirement, and choosing subjectively suitable images for a particular purpose (from a shortlisted set). As expected, we found that similarity-based arrangements are generally more helpful than random arrangements, especially when the user already has some idea of the type of image she is looking for.

Images are used in many different application domains; the ones we chose to study were stock photography and personal photography. We investigated the organisation and browsing of personal photographs in some depth, because of the inevitable future growth in usage of digital cameras, and a lack of previous research in this area.

Full text

PDF (8.3 MB)

BibTeX record

@TechReport{UCAM-CL-TR-543,
  author =	 {Rodden, Kerry},
  title = 	 {{Evaluating similarity-based visualisations as interfaces
         	   for image browsing}},
  year = 	 2002,
  month = 	 sep,
  url = 	 {http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-543.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  number = 	 {UCAM-CL-TR-543}
}