Steady State Random Walks for Path Estimation

摘要

This paper describes a graph-spectral method for path estimation. Our aim is to find a maximum probability path through a lattice of pixel sites. We characterise the path recovery problem using a site transition matrix. A graph-spectral analysis of the transition matrix reveals how the maximum probability path can be located using an eigenvector of the associated normalised affinity matrix. We demonstrate the utility of the resulting method on the problem of recovering surface height from a field of surface normals.