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 language||American English|
|Title of host publication||Computer Vision – ECCV 1996 - 4th European Conference on Computer Vision, Proceedings|
|Editors||Bernard Buxton, Roberto Cipolla|
|Number of pages||15|
|ISBN (Print)||3540611231, 9783540611233|
|State||Published - 1996|
|Event||4th European Conference on Computer Vision, ECCV 1996 - Cambridge, United Kingdom|
Duration: 15 Apr 1996 → 18 Apr 1996
|Name||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Conference||4th European Conference on Computer Vision, ECCV 1996|
|Period||15/04/96 → 18/04/96|
Bibliographical notePublisher Copyright:
© Springer-Verlag Berlin Heidelberg 1996.