Cornell University
Light and Motion: Modeling and Visualizing How Scenes Change Over Time
Abstract
dc:description.abstractThe visible world changes in many ways as time passes. Image and video data capturing these changes is plentiful, but analyzing such data presents many challenges. This thesis presents a set of techniques for modeling, analyzing, and visualizing two key aspects of visual change over time: illumination and motion. The first part discusses the challenging problem of illumination modeling in unstructured photo collections, proposing an approach for shadow detection and sun direction estimation that works on untagged photo collections downloaded from the internet. The second part of the thesis begins by presenting a state-of-the-art method for using motion cues to segment moving foreground objects from static backgrounds. Then, we pose the new problem of how to visualize motions at multiple timescales---from seconds to minutes to hours---in fixed-viewpoint timelapse videos. Two approaches to solving this problem are presented, one based on optimization-driven motion normalization and the other building on the motion segmentation algorithm to segment different timescale videos into foreground layers. The thesis concludes with a discussion of future work, including how the study of illumination and motion might come together to enable further progress in both computer vision and visualization applications.
Degree
thesis:*- Name thesis:degree_name
- Ph. D., Computer Science
- Level thesis:degree_level
- Doctor of Philosophy
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Cornell University
- Year dc:date.issued
- 2018
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wehrwein, Scott
- Chair dc:contributor.chair
-
- Snavely, Keith Noah
- Committee members dc:contributor.committeemember
-
- Bala, Kavita
- Van Loan, Charles Francis
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Dc Identifier Bibid
- 10489835
- Dc Identifier Other
-
ProQuest Submission ID: 10900
ProQuest Publication ID: 10828234