Recent Development in Tensor Voting
Abstract
In this 20 min presentation, I'll talk about what we've
progressed from the 2D and 3D Tensor Voting. Surface
curvature gives a unique, viewpoint independent description
for local shape. In our previous versions, we did not
estimate or use this information explicitly for the shape
inference process. Here, through the use of homogeneous
coordinates and curvature votes, we develop a robust,
voting-based method that both estimates and incorporates
curvature information into the current voting framework.
In contrast to most previous approaches, No partial
derivatives or other second order derivatives are computed,
which are very extremely unstable. On the other hand, the
tensor voting formalism has also been extended into higher
dimensions. We have also applied the 8D version in
robust epipolar geometry estimation that infers the
hyperplane containing all the correct matches given a
set of noisy point correspondences from two unregistered
images, even in the presence of moving objects. Despite
its simiarity to Hough Transform, by use of adequate
data structures and locality, we can show that the space
and time complexity of our method are independent of the
dimensionality. Regardless of the dimensionality, our
experiments show the different versions of our methods
still work and degrades very gracefully in the presence
of large amount of outlier noise (for example, only 1
out of 4 data points is good, i.e. the majority of the
data is noise.)
On-line references
Integrated Surface Inference from Sparse Data Sets