Abstract
This paper addresses the fundamental problem of computing stable medial representations of 3D shapes. We propose a spatially adaptive classification of geometric features that yields a robust algorithm for generating medial representations at different levels of abstraction. The recently introduced continuous scale axis transform serves as the mathematical foundation of our algorithm. We show how geometric and topological properties of the continuous setting carry over to discrete shape representations. Our method combines scaling operations of medial balls for geometric simplification with filtrations of the medial axis and provably good conversion steps to and from union of balls, to enable efficient processing of a wide variety shape representations including polygon meshes, 3D images, implicit surfaces, and point clouds. We demonstrate the robustness and versatility of our algorithm with an extensive validation on hundreds of shapes including complex geometries consisting of millions of triangles.
Introduction: The Instability of Medial Axis Transforms
In geometric modeling, the Medial Axis Transform (MAT) is a powerful tool for representing 3D shapes. Conceptually, it is the set of centers of all maximal inscribed balls within a shape. While theoretically elegant, the MAT is notoriously unstable in practice. A microscopic perturbation on the surface of a 3D model can cause a disproportionately large, jagged branch to sprout from the skeleton.
Traditional "pruning" methods attempt to fix this by removing branches based on local geometric criteria, such as the angle between contact points. However, these local approaches often fail when a shape possesses features at multiple scales. For instance, pruning enough to remove noise from a large body might inadvertently delete small but significant features like teeth or fingers.
Methodology: The Discrete Scale Axis Transform
The authors bridge the gap between the theoretical Continuous Scale Axis Transform and discrete 3D data through a four-stage pipeline:
- Conversion to Union of Balls: The input shape is first approximated as a finite union of "polar balls" derived from the Voronoi diagram of the surface.
- Multiplicative Scaling: This core innovation involves multiplying the radius of every medial ball by a scale factor $s \ge 1$. As balls expand, larger balls "swallow" smaller nearby balls that represent noise or insignificant features.
- Medial Axis Re-computation: The exact medial axis is computed for this newly "grown" union of balls. Because noise-induced balls were absorbed, the resulting skeleton is significantly cleaner.
- Shrinking and Filtering: The balls are returned to their original relative size. A topology-preserving "angle filtration" is then applied to remove minor discretization artifacts without compromising the overall structure.
Key Insights and Results
The "aha moment" of this research is that feature importance should be determined by global context rather than local heuristics. By scaling balls, the algorithm allows the overall shape to dictate which features are preserved. A small feature isolated from others survives, while a similar feature near a massive one is absorbed.
The parameter $s$ serves as a "zoom" for abstraction. At $s=1.0$, the result is the full, noisy medial axis. As $s$ increases, the skeleton simplifies into a stable representation of the object's core topology.
Validation and Performance
The algorithm was validated against the Chen et al. [2009] database of 380 models, including CAD parts and organic shapes, using consistent parameters. It successfully processed high-complexity meshes, such as a Raptor model with 1.7 million triangles.
Computationally, the "Discrete Scale Axis" step is efficient. For a 177k-triangle model, the core computation took just 2.14 seconds, although the initial conversion to balls remains a bottleneck (taking approximately 16 minutes).
Conclusion and Future Work
This approach provides a robust, mathematically grounded method for computing stable skeletons across diverse 3D formats. While highly effective, the method has limitations: it cannot perfectly represent sharp convex features (like cube edges) using a finite union of balls, and extreme abstraction levels can occasionally produce "unnatural" branch merges. Future work may focus on optimizing the initial sampling stage to further reduce total processing time.
Summary updated