3D Dynamic Scene Reconstruction from Multi-View Image Sequences
Speaker: Carlos Leung, ITEE
When: 2003-04-08 11:00:00
Venue: 78-622
Host: A/Prof Brian Lovell
Abstract:The task of computer vision and image processing is to be able to
bring sight to the computer and provide it with vision analysis.
Being able to restore the depth information of an image and recreate
the original 3D scene from images alone have many applications in
computer vision. The reconstruction of a dynamic, complex 3D scene
from multiple images has been an old and challenging problem. While
numerous studies have been conducted on various aspects of this
general problem, such as the recovery of the epipolar geometry
between two stereo images, the calibration of multiple camera views,
stereo reconstruction by solving the correspondence problem, the
modelling of occlusions, and the fusion of stereo and motion, little
has yet been done to produce a unified framework to solve the
general reconstruction problem.
In recent years, a new method for 3D reconstruction has been
introduced. Rather than solving the typical correspondence problem
in stereo analysis, reconstruction is achieved through volumetric
scene modelling. Kutulakos and Seitz introduced the Space Carving
Algorithm aimed at solving the N-view shape recovery problem. The
computed photo hull is determined by computing the photo-consistency
of each voxel through projections onto each available image.
However, while these approaches have produced excellent outcomes,
apart from the fact that they require a vast number of input images,
improvements can still be made by imposing spatial coherence,
replacing the voxel-based analysis with a surface orientated
technique. The goal of this work is to research into more accurate
and efficient methods to reconstruct 3D scenes by developing new
approaches which combines the advantages of surface evolution and
volumetric scene modelling techniques.
Furthermore, while numerous studies have been conducted to
reconstruct 3D scenes from multiple images, little work has been
done in dynamic 3D scene reconstruction from a set of multiple view
image sequences. We propose an alternative view to the stereo and
motion analysis of stereo rigs, by introducing the design of a 4D
voxel volume which aims to recover the 3D structure and the motion
in the scene through volumetric modelling techniques. By augmenting
the design of our 3D voxel feature volume to a 4-D feature space, a
novel formulation of the stereo motion problem is proposed for the
analysis of multi-view image sequences.
Biography:Carlos Leung received his bachelors degree in Electrical Engineering
from the University of Queensland in 2001. He is currently
undertaking a PhD in the field of Computer Vision and Image Analysis
at the University of Queensland. His research interests include 3D
reconstruction and stereo motion.
Type: Ph.D confirmation
Contact:A/Prof Brian Lovell, seminar host (lovell@itee.uq.edu.au)
or Guido Governatori (ITEE seminar co-ordinator)
(guido@itee.uq.edu.au)