Measures for silhouettes resemblance and representative silhouettes of curved objects

Yoram Gdalyahu, Daphna Weinshall

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

6 Scopus citations


We claim that the task of object recognition necessitates a measure for image likelihood, that is: a measure for the probability that a given image is obtained from a familiar (pre-investigated) object. Moreover, in a system where objects are represented by 2D images, the best performance is achieved if those images are selected according to a maximum likelihood principle. This is equivalent to maximum stability of the image, or minimal change under a viewpoint perturbation. All of those qualities involve a quantitative comparison of similarity between images. We propose different metric functions which can be imposed on the image space of curved three dimensional objects. We use these metrics to detect the representative views (most stable and most likely views) of three test models. We find the same representative views under all the investigated metrics, suggesting that local maxima of stability and likelihood are metric independent. Our method of image comparison is based solely on the appearance of the occluding contour, hence our method is suitable for object recognition from silhouettes.

Original languageAmerican English
Title of host publicationComputer Vision – ECCV 1996 - 4th European Conference on Computer Vision, Proceedings
EditorsBernard Buxton, Roberto Cipolla
PublisherSpringer Verlag
Number of pages15
ISBN (Print)3540611231, 9783540611233
StatePublished - 1996
Event4th European Conference on Computer Vision, ECCV 1996 - Cambridge, United Kingdom
Duration: 15 Apr 199618 Apr 1996

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference4th European Conference on Computer Vision, ECCV 1996
Country/TerritoryUnited Kingdom

Bibliographical note

Publisher Copyright:
© Springer-Verlag Berlin Heidelberg 1996.


Dive into the research topics of 'Measures for silhouettes resemblance and representative silhouettes of curved objects'. Together they form a unique fingerprint.

Cite this