Learning to Segment Images Into Material and Object Classes
McHenry, Kenton Guadron
This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.
Permalink
https://hdl.handle.net/2142/81819
Description
Title
Learning to Segment Images Into Material and Object Classes
Author(s)
McHenry, Kenton Guadron
Issue Date
2008
Doctoral Committee Chair(s)
Ponce, Jean
Department of Study
Computer Science
Discipline
Computer Science
Degree Granting Institution
University of Illinois at Urbana-Champaign
Degree Name
Ph.D.
Degree Level
Dissertation
Keyword(s)
Computer Science
Language
eng
Abstract
This dissertation addresses the task of learning to segment images into meaningful material and object categories. With regards to materials we consider the difficult task of segmenting objects made of transparent materials such as glass. To do this we consider information in the form binary features. Unlike more traditional unary features which consider information contained within a single location, binary features which consider information between pairs of locations are used to capture the notion of transparency (i.e. being able to see through something). We begin by using this in an edge based approach to locate the edges of glass objects. Segmenting transparent regions, which is desirable in order to locate objects, is ambiguous with this binary information alone. We deal with this by treating this information as a measure of discrepancy, relating how different two regions are from one another. We then combine this with a complimentary affinity measure which relates how well two regions belong together. These two measures are then combined within a single energy function which can be optimized to segment regions of transparent material. With regards to opaque objects an initial segmentation can be constructed using local features within regions produced from an over segmentation of the image. Our interest here is in improving these local segmentations by incorporating global information. Using global features (i.e. features that consider all regions simultaneously) and synthetically generated contrastive data an energy based model is constructed to estimate the quality of a given segmentation. Based on these segmentation quality estimates we attempt to improve a given segmentation.
Use this login method if you
don't
have an
@illinois.edu
email address.
(Oops, I do have one)
IDEALS migrated to a new platform on June 23, 2022. If you created
your account prior to this date, you will have to reset your password
using the forgot-password link below.
