A Computer Vision Approach for Visualization from Sparse, Noisy 3-D Data
We are interested in feature extraction from volume data
in terms of coherent surfaces and 3-D space curves. The
input can be an inaccurate scalar, vector or tensor field,
sampled densely or sparsely on a regular 3-D grid.
Even if such data is obtained from state-of-the-art
sensors or data sources, unavoidable outlier noise and missing
data can make the traditional iso-surface techniques
inappropriate. Here, we propose a unified computational
scheme, Tensor Voting, for dealing with such real data.
The representation scheme is one of a second-order symmetric
tensor that encodes preferred direction, as well as orientation
uncertainty information at discontinuities. Data communication
is achieved by a novel voting process that simultaneously infers
smooth structures, detect orientation discontinuities, and ignores
outlier noise. This process produces a dense tensor map, which
is further decomposed into a surface, a curve, and a junction
map. A modified marching process is then used for coherent
surface and curve extraction. In this talk, a live demo will
be given, and other application domains such as
flow visualization, vortex extraction, terrain, seismic data processing,
and some medical imagery applications will be discussed.
Integrated Surface Inference from Sparse Data Sets