| The theme of this thesis is to complete the 3D tensor voting theory for computer vision and graphics applications by incorporation of (i) tensor of curvature estimation, (ii) first order voting, (iii) automatic multiple scale data handling, and deriving pertinent applications with better results. The original tensor voting is a unified, noniterative and robust framework for feature extraction and reconstruction. Although it is promising in many computer vision applications, a lot of issues still existed. The estimation of curvature information, such as maximum and minimum principal curvature magnitudes and directions, is absent in the original framework. In this thesis, a robust method for estimation of the curvature tensor is first designed to address this issue. The inference of first order features such as region boundaries, surface boundaries and curve endpoints, is also missing in the original 3D framework. This issue is also addressed, making more applications possible. The original framework also lacks a good foundation for multiscale and multiresolution data processing. Inconsistent feature extraction might arise when the input data consists features of various scales. Therefore, an automatic multiscale feature extraction for tensor voting is proposed and implemented. One practical application of the new framework which is impossible for the original framework is the reconstruction and visualization for huge, dense or quasi-dense data. A new approach for on-demand reconstruction based on tensor voting is proposed. |
