Skeleton-Space: a Multiscale Shape Description Combining Region and Boundary Information

Robert L. Ogniewicz

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Pages 746-751, 1994


Motivation

The medial axis transform (MAT) is a frequently used technique for producing a skeleton-like description of planar shapes called the medial axis. Unfortunately, the MAT overemphasizes boundary details (e.g. noise on the boundary), making it difficult to match between medial axis components at different levels of detail. Also certain 2D operations such as blurring will even more significantly alter the structure of the medial axis than boundary smoothing.

Goal of This Research

The authors wish to design a multiscale extension of the MAT which is not overly sensitive to small variations on the boundary or to 2D operations that do not significantly change the shape of an object.

Goal of This Paper

This paper presents a new technique to compute a multiscale MAT which is more robust with respect to small variations in shape. The method starts out with the MAT of the original, nonsmoothed object. A selection scheme is used to organize the components of the medial axis according to their structural importance. The multiscale MAT combines boundary properties from distinct scales with the regional information of the medial axis, unlike many standard simplification methods that are solely boundary-based or region-based.

Related Work

Results

Advantages of the multiscale MAT proposed in this paper include:
  1. No correspondence problem must be solved between distinct scales
  2. The technique is much less sensitive to small boundary variations than many existing methods
  3. Coarse scale discontinuities along the boundary are largely retained
  4. The hierarchical MAT allows automatic selection of feasible scale and regularization parameters
  5. The technique is very time-efficient since the MAT need not be reevaluated for each scale

Bibliography


ctj