hide
Free keywords:
-
Abstract:
As computers can only represent and process discrete data, information gathered
from the real world always has to be sampled. While it is nowadays possible to
sample many signals accurately and thus generate high-quality reconstructions
(for example of images and audio data), accurately and densely sampling 3D
geometry is still a challenge. The signal samples may be corrupted by noise and
outliers, and contain large holes due to occlusions. These issues become even
more pronounced when also considering the temporal domain. Because of this,
developing methods for accurate reconstruction of shapes from a sparse set of
discrete data is an important aspect of the computer graphics processing
pipeline.
In this thesis we propose novel approaches to including semantic
knowledge into reconstruction processes using template based shape processing.
We formulate shape reconstruction as a deformable template fitting process,
where we try to fit a given template model to the sampled data. This approach
allows us to present novel solutions to several fundamental problems in the
area of shape reconstruction. We address static problems like constrained
texture mapping and semantically meaningful hole-filling in surface
reconstruction from 3D scans, temporal problems such as mesh based performance
capture, and finally dynamic problems like the estimation of physically based
material parameters of animated templates.
