Medial Axis Computation and Evolution

Sylvain Bouix, Kaleem Siddiqi, Allen R. Tannenbaum, Steven W. ZuckerStatistics and Analysis of Shapes 2006 · DOI · DBLP

Abstract

The medial axis provides a compact representation of shape topology and geometry, yet its computation is notoriously sensitive to boundary noise. This work introduces a robust framework based on the Average Outward Flux (AOF) of the distance gradient. By leveraging integral measures and deriving the mathematical evolution of the skeleton, the authors demonstrate a stable method for shape partitioning and analysis in both 2D and 3D.

Introduction

The medial axis of a shape—defined as the set of centers of maximal inscribed disks—is a fundamental tool in computer vision and medical imaging. It simplifies complex shapes into a "skeleton" that describes topology, thickness, and orientation. However, the medial axis is notoriously "brittle": small perturbations on the boundary can lead to large, spurious skeletal branches that hinder practical analysis. This paper presents a robust computational approach and a formal mathematical analysis of how these skeletons evolve as a shape changes.

Robust Computation via Average Outward Flux

Traditional skeletonization methods often rely on local differential properties, such as gradients, which tend to amplify noise. To address this, the authors propose a Hamilton-Jacobi formulation using the Average Outward Flux (AOF) of the distance function's gradient.

As an integral measure, the AOF sums information over a local region, making it inherently more robust than local calculations. A key theoretical contribution is the link between AOF and the object angle—the angle between the inward normals at the points where a maximal disk touches the boundary. The research demonstrates that skeletal sensitivity is inversely proportional to the sine of this angle. Consequently, the AOF provides a natural "salience" measure that distinguishes true skeletal features from noisy artifacts.

Topology-Preserving Thinning and Evolution

The framework employs a topology-preserving thinning algorithm guided by the AOF. This algorithm removes pixels or voxels from the outside in, ensuring that the object's topology (e.g., holes and connectivity) remains intact. This process effectively turns a raw voxel cloud into a structured representation of the shape.

Furthermore, the authors derive closed-form equations for the velocity of medial axis points. These equations describe exactly how the skeleton moves in response to boundary evolution. For instance, when a boundary is smoothed via a geometric heat equation, the medial axis evolves according to a related smoothing flow. This provides a rigorous theoretical foundation for "pruning" techniques, where boundary smoothing is used to simplify the resulting skeleton.

Results and Applications

The methodology was validated on several datasets, including 2D silhouettes of animals (such as horses and elephants) and 3D human models in various poses (running, jumping). The AOF-based thinning successfully extracted clean, well-connected skeletons even for complex, articulated 3D shapes.

This approach allowed for the automatic partitioning of shapes into meaningful components, such as limbs and torsos. By labeling skeletal points as surfaces, curves, or junctions, the method transforms raw data into a structured "hypergraph" representation. While the evolution equations are powerful, the authors note they require specific conditions on boundary speed to remain valid, suggesting that future research into "medial density" could further refine how we quantify the importance of different skeletal branches.

Conclusion

By shifting from differential to integral measures, this work provides a theoretically grounded and practically robust approach to medial axis computation. The discovery of the relationship between AOF and the object angle, combined with the derivation of skeletal evolution equations, bridges the gap between boundary geometry and internal skeletal structure, offering a powerful framework for shape analysis.

Summary updated

BibTeX
@inproceedings{bouix2006medial,
  author = {Sylvain Bouix and Kaleem Siddiqi and Allen R. Tannenbaum and Steven W. Zucker and Allen Tannenbaum and Allen Tannenbaum and Steven W. Zucker},
  title = {{Medial Axis Computation and Evolution}},
  booktitle = {Statistics and Analysis of Shapes},
  year = {2006},
  doi = {10.1007/0-8176-4481-4\_1},
}
Figure