Advances in 3D Shape Acquisition (thesis)
In this dissertation we discuss a variety of techniques that
advance the state of the art in the field of 3D shape
acquisition from real world objects. The research was done
in collaboration with Szymon Rusinkiewicz, James Davis, Ravi
Ramammorthi, and Tim Weyrich.
Our first contribution is a new framework for the
classification of stereo triangulation algorithms. We
classify methods according to the dimensions along which
observations by both cameras are matched against each other.
Different algorithms consider information that extends in
space, in time, or simultaneously in both dimensions. Based
on this framework, we design a novel algorithm for the
triangulation of dynamic objects, as well as a new stereo
setup based on unstructured active lighting.
We then present a novel sub-pixel precision refinement
algorithm for stereo matches. We treat both cameras
symmetrically, instead of assuming one camera to provide a
reference image to be matched against. By refining match
coordinates simultaneously on both cameras, we avoid a
source of bias that can otherwise manifest itself as
coherent noise in the reconstructions.
We also provide an efficient algorithm for combining
position and orientation measurements into an optimal
surface. Since position and orientation measurements are
obtained from independent sources, each contains errors with
distinct frequency characteristics. By optimizing a surface
to conform to the most precise frequency components from
each source, we can produce reconstructions that are
substantially more precise than the original measurements.
Finally, we present a strategy for the acquisition of the 3D
shape of shiny objects. Standard triangulation strategies
that rely on captured appearances fail due to the view
dependent nature of the images of such objects. We present a
matching cost function based on surface normal consistency
that can be used with standard dense stereo matching
algorithms, and discuss the ambiguities that can arise.