Files in this item



application/pdfSITTIG-THESIS-2015.pdf (4MB)
(no description provided)PDF


Title:Enhanced image-based 3D reconstruction with 2D-to-3D constraints
Author(s):Sittig, Michael Glenn
Advisor(s):Golparvar Fard, Mani
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Structure from Motion (sfm)
mesh assisted structure from motion (mesh assisted sfm)
image based 3d reconstruction
Abstract:This paper discusses an improvement to the Mesh-Assisted Structure from Motion algorithm. Mesh-Assisted Structure from Motion is an improvement over the original Structure from Motion algorithm, which generates a 3D reconstruction of a scene based on a set of 2D images depicting the scene. Mesh-Assisted Structure from Motion improves on it by incorporating a mesh. A user registers one of the images in the image set to the mesh and provides this registration as a constraint to the Structure from Motion pipeline. This solves some difficulties the default Structure from Motion algorithm has, and results in higher reconstruction quality. The improvement discussed in this paper allows additional anchor images to be registered to the mesh and provided as constraints to the pipeline. Using additional anchor images results in fewer images needing to be registered during pipeline execution. This means that more of the pipeline may be executed automatically, with no additional user input. Depending on the available images and mesh, using additional anchor cameras may or may not improve reconstruction quality. For most data sets, additional anchor images result in a slightly worse reconstruction. However, for some data sets, the reconstruction quality may be improved by using more anchor images. The most important criteria for determining if using more anchor images will improve or harm the reconstruction accuracy is the accuracy of the input mesh. However, even when the reconstruction becomes less accurate, using more anchor images will still decrease the amount of input necessary during execution, and thus allow more of the pipeline to be executed without needing to wait for user input. Finally, several possible improvements to the modified Mesh-Assisted Structure from Motion algorithm are discussed. These improvements mostly relate to improving the anchor image selection process, to choose images which will result in a more accurate reconstruction. These improvements would increase the number of situations in which using the modified algorithm with multiple anchor images is superior to the algorithm with only a single anchor image.
Issue Date:2015-12-11
Rights Information:Copyright 2015 Michael Sittig
Date Available in IDEALS:2016-03-02
Date Deposited:2015-12

This item appears in the following Collection(s)

Item Statistics