Stabilizing Medial Axis Transforms Against Boundary Noise

92 concepts74 papers1 proposal

The Medial Axis Transform (MAT)—often called a geometric skeleton—is one of the most elegant concepts in shape analysis. Imagine setting fire simultaneously to the entire boundary of a shape. As the fire burns inward at a uniform speed, the locations where the fire fronts collide form the "skeleton" of the shape.

By reducing a complex 2D silhouette or 3D volume to a 1D graph of thin curves and junctions, skeletonization provides a compact topological backbone. It is widely used across computer vision, from robotic mapping (Arno Bücken et al.) to medical image analysis (Enzo Ferrante et al.).

Yet, traditional skeletonization possesses a fatal flaw: extreme sensitivity to boundary noise. A microscopic perturbation on an object's surface can radically alter the global topology of its skeleton, rendering standard algorithms brittle in real-world applications. Over the past few decades, researchers have developed a variety of mathematical frameworks to resolve this instability.


The Intuition of Instability: Why Skeletons Shatter

To understand why skeletons are so sensitive, return to the fire front analogy. If an object has a perfectly smooth boundary, the fire fronts meet along a clean, central spine.

Now, introduce a tiny, microscopic bump—a single noisy pixel—to the boundary. This bump acts as an independent, premature ignition point. The fire front expanding from this tiny bump will quickly collide with the main advancing fire front. Because this collision happens deep within the shape's interior, it projects a brand-new, long skeletal branch (often called a "spur," "spike," or "parasitic branch") all the way back to the bump.

As noted in the benchmark study SkelNetOn 2019: Dataset and Challenge on Deep Learning for Geometric Shape Understanding, minor surface irregularities inevitably generate significant, erroneous skeletal branches when using purely geometric algorithms.

Smooth Boundary:          Noisy Boundary:
   ___________               _____/\_____
  |           |             |            |
  |  -------  |             |  ---|----  |  <-- Spurious branch 
  |___________|             |_____|______|      triggered by a tiny bump

This is not just a computational nuisance; it is a topological catastrophe. In graph-based shape matching, a single spurious branch introduces false nodes and edges, completely changing the graph's structure. As Longin Jan Latecki et al. and Andrea Torsello et al. demonstrate, this structural volatility makes raw digital skeletons highly unreliable for consistent object recognition or shape retrieval under slight rotations, scaling, or discretization noise.


The Mathematical Root Cause

Why do standard computational approaches—such as Voronoi diagrams, digital thinning, or distance transforms—fail so consistently in the presence of noise?

The problem lies in their reliance on local derivatives. Most skeletonization algorithms locate the medial axis by calculating the gradient of a Euclidean distance field (the distance from any interior point to the closest boundary point). At the medial axis, these gradients become discontinuous because multiple boundary points are equidistant.

Pavel Dimitrov explains that because these local derivatives look for exact symmetry, even a minor boundary ripple drastically flips the direction of the local gradient vectors. In a foundational paper on parallel thinning, Antoine Manzanera identified a fundamental theoretical incompatibility: it is mathematically impossible to strictly adhere to the exact geometric center of a digital shape while remaining robust against image noise.


Strategies for Topological Stability

To overcome this limitation, researchers have shifted away from strict local symmetry toward global, hierarchical, and optimization-driven formulations.

1. Global Flux Fields Instead of Local Derivatives

Instead of analyzing local distance gradients, some frameworks treat the distance field as a physical fluid flow and calculate its net outward flux.

Pavel Dimitrov and Morteza Rezanejad et al. demonstrate that by integrating the gradient vectors over a local region, high-frequency boundary noise naturally cancels out. The global integration smooths over minor surface irregularities, ensuring that a skeletal branch is only generated if there is a substantial, macro-scale convergence of the shape's geometry.

2. Hierarchical Abstraction and "Swallowing" Mechanisms

Rather than preventing noisy branches from forming, another paradigm allows them to generate but manages them through structural hierarchies.

Diego Macrini et al. introduced Bone Graphs, which categorize the medial axis into stable, primary components ("bones") and transient, noise-driven components ("ligaments"). This abstraction insulates the core topology from boundary noise.

For 3D shapes, Cheng Lin et al. proposed a geometric swallowing mechanism. During the skeletonization process, major skeletal branches actively absorb unstable, minor spikes. This prevents over-segmentation and ensures the final structural decomposition remains stable even when the underlying mesh is highly irregular.

3. Reformulating Skeletonization as a Set Cover Problem

Instead of thinning a shape from the outside in, what if we fill it from the inside out?

Zhiyang Dou et al. pioneered this perspective with the Coverage Axis. They framed skeletonization as a optimization problem: find the minimum number of interior "medial balls" required to fully cover the object's surface. Because this method optimizes for global volume coverage rather than exact local symmetry, it ignores superficial surface perturbations, yielding clean, stable skeletons from messy, non-watertight real-world scans.

4. Disconnected Structural Representations

Traditional skeletal graphs are fragile because a single noisy branch alters the connectivity of the entire network. To bypass this entirely, Cagri Aslan et al. proposed an alternative: representing objects as a collection of disconnected skeletal branches rather than a unified graph.

By measuring the spatial arrangement of these disconnected segments within a shape-centered coordinate system, local boundary noise is completely isolated. A spur on the "hand" of a shape might alter one local branch, but it can no longer trigger a global topological collapse across the rest of the model.

Key Foundational Work

Learning Topological Maps: An Alternative Approach (1996) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 67 papers
  2. Version 1 · 2026-10-04 · 99 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_boundary_noise_sensitivity_8d0bfceb,
  title = {boundary noise sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/boundary-noise-sensitivity/rdr_8d0bfcebaca847acbaeef1a33c9c756b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_boundary_noise_sensitivity_8d0bfceb_v2,
  title = {boundary noise sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/boundary-noise-sensitivity/rdr_8d0bfcebaca847acbaeef1a33c9c756b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake