Segmentation of 3D Images Using Gaussian and Mean Curvatures Sign Classification and Relaxation Labelling Optimization

摘要

The paper describes how range images can be segmented using the sign of differential geometry operators such as the mean and Gaussian curvatures. This segmentation called topographic primal sketch is invariant to rigid body motion in space and is defined by eight fundamental surfaces. The first part of the paper presents two novel techniques where an initial estimate of the categories, full of inconsistent labelling due to noise, is transformed into a consistent one. One of the two methods is based on a label relaxation technique, where consistency is viewed as a local optimization problem, and the second is based on stochastic relaxation, where the local classification of a pixel is solved by statistical decisions. The advantages and disadvantages of each method are presented.