In this paper, a general technique is proposed for the analysis of multi-dimensional feature space. The basic computational module of the technique is the tensor voting theory, which was formerly used for structure inference from sparse data. We analyze the methodology of tensor voting 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. Algorithms for two low-level vision tasks, discontinuity preserving smoothing and motion layer inference, are described as applications of tensor voting. Several experimental results illustrate its excellent performance.