Please use this identifier to cite or link to this item: http://scholarbank.nus.edu.sg/handle/10635/20435
Title: A Fixation Based Segmentation Framework
Authors: AJAY KUMAR MISHRA
Keywords: segmentation, fixation, visual attention, moving object segmentation, cue combination, shape based segmentation
Issue Date: 2-Jun-2010
Source: AJAY KUMAR MISHRA (2010-06-02). A Fixation Based Segmentation Framework. ScholarBank@NUS Repository.
Abstract: This thesis proposes a novel fixation based segmentation algorithm that takes a point (``fixation'') in the image (or video) as input and outputs the region containing that fixation point. Such formulation of segmentation with fixation is a well-defined problem in contrast with the traditional segmentation formulation that tries to break an image (or a scene) into mutually exclusive regions. The is shown qualitatively in the thesis that the traditional definition of segmentation is not well defined and that segmentation can be defined optimally only if the object of interest is identified prior to segmentation. The proposed algorithm carries out segmentation as a two-step process. In the first step, all visual cues, static monocular, stereo and(or) motion, are used to generate a probabilistic boundary edge map that contains the probability of an edge pixel being at a depth boundary. In the second step, the probabilistic boundary edge map is transformed from the Cartesian space to the polar space with the fixation point chosen as the pole for this transformation. In the polar space, the segmentation of the fixated region is carried out as a binary labeling problem that finds the optimal cut through the polar edge map, which becomes the closed contour around the fixation point as it gets transformed back to the Cartesian space. Motivated by the experiments in the psychophysics that suggest humans do not just make a single fixation but a series of them, a subset of which are related to each other, we propose a multi-fixation strategy that starts with a given fixation and makes a series of dependent fixations to segment the object of interest completely even when the shape of the object is complex. The multiple fixation strategy is also used to segment complex shaped objects especially the ones with a thin elongated part. Finally, an attempt to use the sparse motion information instead of the dense optic flow map to segment moving objects has also been made. The motivation for this lies in the difference between motion and color cues. Motion cues are inherently sparse since motion can be detected unambiguously only at some salient locations in the scene whereas color cues are known at every location with high accuracy. We propose an algorithm to segment moving objects in a video without having to calculate the dense optic flow map of the scene.
URI: http://scholarbank.nus.edu.sg/handle/10635/20435
Appears in Collections:Ph.D Theses (Open)

Show full item record
Files in This Item:
File Description SizeFormatAccess SettingsVersion 
MishraAK.pdf6.41 MBAdobe PDF

OPEN

NoneView/Download

Page view(s)

302
checked on Dec 11, 2017

Download(s)

186
checked on Dec 11, 2017

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.