Abstract:
In this paper we describe a graph-spectral method for 3D surface segmentation from 2D imagery. The method locates patches by finding groups of pixels that can be connecte...Show MoreMetadata
Abstract:
In this paper we describe a graph-spectral method for 3D surface segmentation from 2D imagery. The method locates patches by finding groups of pixels that can be connected using a curvature minimising path. The path is the steady state Markov chain on transition probability matrix. We provide two methods for computing this matrix. The first uses the information provided by the field of surface normals extracted from the 2D intensity image using shape-from-shading. Here we compute the elements of the transition matrix using the change in surface normal directions to estimate the normal curvature. The second approach uses the raw image brightness together with a Lambertian reflectance model to make estimates of curvature. We compare the surface segmentations delivered by these two methods with those obtained using shape-index maximal patches.
Published in: 2002 International Conference on Pattern Recognition
Date of Conference: 11-15 August 2002
Date Added to IEEE Xplore: 10 December 2002
Print ISBN:0-7695-1695-X
Print ISSN: 1051-4651