Feature space analysis using low-order tensor voting

In this paper, low-order Tensor Voting, which was formerly used for structure inference from sparse data, is extended for feature space analysis. It is a nonparametric technique, because it does not have embedded assumptions. The methodology and possible applications are analyzed systematically. Its relation to Kernel Density Estimation and Mean Shift is also established, based on what the utilities for two fundamental analyses of feature space, density estimation and mode detection, are discussed. At last, two low-level vision tasks, image segmentation and motion analysis, are described as applications of the low-order Tensor Voting. Several experimental results illustrate its excellent performance.