The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


"THREE-DIMENSIONAL RECONSTRUCTION FROM IMAGES"

By

Mr. Gang Zeng


Abstract

Three-dimensional (3D) reconstruction from images is an old, fundamental
yet difficult problem in computer vision, which has been extensively
investigated over the past three decades. Recent developments in camera
calibration and multimedia computing have broadened the interests in the
reconstruction problem, and the big proliferation of digital cameras and
computers allow to take and process multiple images respectively. The
emphasis of the nowadays applications has been shifted to generating more
"appearance" views of the scene, and it requires highly detailed shapes to
be constructed within a tolerable computational resources. Such a task,
however, is difficult to accomplish due to the intrinsic ill-posedness of
the reconstruction, interference from image noise, and insufficiency of
efficiency and scalability.

Mathematically, scene objects can be represented as smooth surfaces or
curves with arbitrary topology, and three-dimensional reconstruction from
images could be cast into various problems and formulations. In this
dissertation, we focus on the mathematical descriptions and
representations, and discuss the reconstruction approaches that use
volumetric, graph-cut, and level-set optimization tools, and the objective
functionals that use different image cues, such as silhouette, photometry,
texture and shading. The main contribution of this dissertation is a
general introduction, discussion and comparison among different kinds of
image-based reconstruction methods.

Our discussion is accompanied by analysing, designing and implementing of
eight our own reconstruction approaches, covering multi-view volumetric
and surface reconstructions, single-view reconstruction and relighting,
and space curve and branch reconstructions. Meanwhile, these eight
proposed approaches are designed for various setups. Thus, they are
different from each other, and have their own distinct features. More
precisely, Silhouette Carving from Multiple Images of an Unknown
Background constructs the visual hull without knowing silhouette
information; Robust Carving for Non-Lambertian Objects is designed for the
reconstruction of non-Lambertian objects; Progressive Surface
Reconstruction from Images using a Local Prior produces high quality
surfaces by local optimizations; Surface Reconstruction by Propagating 3D
Stereo Data in Multiple 2D Images reconstructs the surface by propagating
from 3D stereo points; Interactive Shape from Shading is an interactive
approach to producing the surface of a single shading image; Single-View
Relighting with Normal Map Painting assigns normals for relighting without
constructing the surface geometry; Carving 3D Edges and Curves from
Multiple Views is a carving approach for space curve reconstruction;
Image-based Plant Modeling produces photo realistic 3D models for
real-world plants. Real scene examples demonstrate the quality,
performance, and usability of these reconstructions for nowaday real
applications.


Date:			Thursday, 24 August 2006

Time:			4:00p.m.-6:00p.m.

Venue:			Room 5501
			Lifts 25-26

Chairman:		Prof. Allen Moy (MATH)

Committee Members:	Prof. Long Quan (Supervisor)
			Prof. Chiew Lan Tai
			Prof. Chi Keung Tang
			Prof. In So Kweon (Elec. Engg., KAIST)
			Prof. Anil Jain (Comp. Sci. & Engg.,
					 Michigan State Univ.)


**** ALL are Welcome ****