An Enhanced Tensor Voting Framework and Applications

Wai-Shun Dickson Tong


Abstract

Recently, a computational framework for feature extraction and segmentation, Tensor Voting, has been proposed. It has demonstrated its efficacy in various vision problems with encouraging results. It is a unified framework, which makes use of the continuity constraint to infer the geometrical features and generate layered description in terms of surfaces, regions, curves, and junctions. The formalism not only is capable of dealing with multiple structures, detecting and localizing discontinuities, it is also very robust against outlier noise. No iterative optimization search is required.

Further research on the formalism leads to a number of improvements that are originally absent in the basic formalism. This thesis consists of theory and application aspects on tensor voting. We enhanced the basic tensor voting theory with first order information. Inference of important geometrical features, such as region and surface boundaries, and curve endpoints can be achieved. These enhancements enable us to derive an adaptive scale detection algorithm.

With the improved formalism, two applications, one in computer graphics and the other in medical surface extraction, are proposed. A novel and effective algorithm on epipolar geometry estimation, which takes the benefit of the latest advancement of the higher dimensional tensor voting formalism is also derived.


Maintained by Philippos Mordohai