Recent models of object recognition have proposed that objects are recognised on the basis of shape features encoded from specific viewpoints of observation. These models have largely failed, however, to specify the particular shape features involved, and thus are unable to predict the situations in which a change in viewpoint will impair recognition (in terms of accuracy and response times). The reason that models have not specified shape features is because the geometry of 3D objects means that hugelyvariable sets of features are possible. In this project, in collaboration with Dr William Hayward (Chinese University of Hong Kong), we investigate the information conveyed by perspective changes in silhouettes. Initially we determine the conditions under which recognition of silhouettes is as good as recognition of normal objects, and then evaluate the manner in which silhouette information is encoded by the visual system and broken into constituent shape features. Then we test the effect of changing particular shape features in the silhouette on object recognition. In addition, we apply principles of perceptual organisation to isolate a small class of shape features in the bounding contour of an object and thereby determine the manner in which part segmentations influence recognition costs. This project will allow computational approaches to object recognition to develop a set of image features with which to properly test their models.
Spehar B Hayward W "Information contained in static and dynamic silhouette views"
Spehar B Hayward W "The role of depth information in recognition of silhouette views of objects"
ARC Large Grant Scheme (2000-2001).
[ home | research interests | publications | lecture notes | grants | collaborators | curriculum vitae | b.spehar@unsw.edu.au ] |