The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #92112

Title:Precategorical Segmentation Using Disparity
Date: 1992
Abstract:Olson and Coombs have briefly described a method for segmenting an image pre-categorically using information about disparity [Olson & Coombs 90]. Low disparity objects are retained whilst high disparity objects are attenuated and such a process is interesting for several reasons. There is evidence that both humans and primates perform disparity based segmentation [Miles et al 90] and the algorithm could also be used as part of a vergence control system or to simplify the correspondence problem. With this for inspiration the project proposes an algorithm and investigates it thoroughly by first making an implementation and then subjecting the system to various experimental tests. Olson and Coombs' results were replicated and it was found that they had made judicious choices of objects, disparities and type of background in order to achieve what they did. The algorithm was found to attenuate objects if their disparity is greater than the angle that they subtend at the eye. For an inter-ocular separation of 30 cm, when viewing object 2 m away, objects within a radius of 20 cm of fixation are retained and any others are attenuated. The algorithm is able to successfully deal with occluding objects, noise and texture; all traditionally hard problems in segmentation. It makes use of a multiresolution technique and is novel in many respects making it of importance for segmentation research in general.

[Search These Pages] [DAI Home Page] [Comment]