Files in this item

FilesDescriptionFormat

application/pdf

application/pdfLIU-DISSERTATION-2016.pdf (21MB)
(no description provided)PDF

Description

Title:Physically based geometry and reflectance recovery from images
Author(s):Liu, Siying
Director of Research:Do, Minh N.
Doctoral Committee Chair(s):Do, Minh N.
Doctoral Committee Member(s):Kamalabadi, Farzad; Hasegawa-Johnson, Mark A.; Hart, John C.
Department / Program:Electrical & Computer Eng
Discipline:Electrical & Computer Engr
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):texture
RGB-D
reflectance
image correspondence
color transfer
Abstract:An image is a projection of the three-dimensional world taken at an instance in space and time. Its formation involves a complex interplay between geometry, illumination and material properties of objects in the scene. Given image data and knowledge of some scene properties, the recovery of the remaining components can be cast as a set of physically based inverse problems. This thesis investigates three inverse problems on the recovery of scene properties and discusses how we can develop appropriate physical constraints and build them into effective algorithms. Firstly, we study the problem of geometry recovery from a single image with repeated texture. Our technique leverages the PatchMatch algorithm to detect and match repeated patterns undergoing geometric transformations. This allows effective enforcement of translational symmetry constraint in the recovery of texture lattice. Secondly, we study the problem of computational relighting using RGB-D data, where the depth data is acquired through a Kinect sensor and is often noisy. We show how the inclusion of noisy depth input helps to resolve ambiguities in the recovery of shape and reflectance in the inverse rendering problem. Our results show that the complementary nature of RGB and depth is highly beneficial for a practical relighting system. Lastly, in the third problem, we exploit the use of geometric constraints relating two views, to address a challenging problem in Internet image matching. Our solution is robust to geometric and photometric distortions over wide baselines. It also accommodates repeated structures that are commonly found in our modern environment. Building on the image correspondence, we also investigate the use of color transfer as an additional global constraint in relating Internet images. It shows promising results in obtaining more accurate and denser correspondence.
Issue Date:2016-11-04
Type:Thesis
URI:http://hdl.handle.net/2142/95307
Rights Information:Copyright 2016 Siying Liu
Date Available in IDEALS:2017-03-01
Date Deposited:2016-12


This item appears in the following Collection(s)

Item Statistics