Robust Skeletonization for Complex Geometric Structures

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 299 papers130 conceptsfrozen 2026-10-04

A skeleton is the topological "stick figure" of a shape. Formally known as the Medial Axis Transform (MAT), it represents an object as the set of centers of all maximal inscribed disks that touch the object's boundary at two or more points. This representation is incredibly powerful because it captures both the topology (how parts are connected) and the geometry (the thickness at every point) of a shape in a highly compressed form.

However, the medial axis has a notorious flaw: it is mathematically unstable. A single-pixel bump on the boundary of a smooth circle can cause a massive, spurious branch to "sprout" from the center of the skeleton. This sensitivity to noise has historically made skeletons difficult to use in real-world applications like medical imaging, robotics, and computer vision.

The Problem of Spurious Branches

The fundamental challenge in skeletonization is distinguishing between meaningful structural features and noise-induced artifacts, often called "spurs." As Jae-Moon Chung and Noboru Ohnishi noted in Chain of Circles for Matching and Recognition of Planar Shapes, traditional medial axis representations are so sensitive to boundary noise that minor surface irregularities can render the resulting skeleton unusable for reliable object recognition.

In 3D models, this problem is amplified. Ting Sun and Guohua Geng explain in A Curve-Skeleton Extraction Algorithm based on Vector Fields for 3D Model Retrieval that geometric irregularities on a model's boundary lead to redundant branches that do not reflect the core topology. Whether it is a hand-drawn sketch, as explored by Jing Zhang in A novel sketch-based 3D model retrieval approach based on skeleton, or a scanned document, the "thinning" process often introduces false intersections and "blobs" that distort the object's true path.

Moving from Geometry to Physics and Optimization

To overcome this instability, researchers have shifted away from purely geometric definitions of the medial axis toward more robust formulations based on physics and numerical optimization.

One influential approach treats skeletonization as a "flux" problem. In Flux Graphs for 2D Shape Analysis, Morteza Rezanejad introduces a saliency measure based on the "uniqueness of inscribed disks." By treating the boundary as a source of a physical field, researchers can identify the "medial" points where the field's flux is most concentrated. This global view is much more resistant to local high-frequency noise than traditional methods that rely on local derivatives. This same principle allows robots to create stable "road maps" of unknown environments, as demonstrated in Robust environment mapping using flux skeletons.

Another strategy is to frame skeletonization as a fitting problem. Daniel Rebain introduced LSMAT Least Squares Medial Axis Transform, which reimagines the skeleton as a robust numerical optimization. By using a least-squares formulation to fit maximally inscribed spheres, the method can produce stable skeletons even from extremely noisy point clouds where traditional algorithms would fail.

Deep Learning and Topological Priors

The rise of deep learning has provided a new way to bypass the fragility of manual thinning algorithms. Instead of using a sequence of geometric rules, researchers train neural networks to predict the skeleton directly from raw data.

In U-Net Based Convolutional Neural Network for Skeleton Extraction, Oleg Panichev and Alona Voloshyna use a weighted focal loss to train a U-Net to identify skeletal pixels, effectively learning to ignore boundary noise that would trip up classical methods. Similarly, Dr-SAM adapts Meta’s Segment Anything Model to extract vascular skeletons from angiography images, combining deep learning's segmentation power with topological constraints to ensure the resulting "stick figure" is anatomically plausible.

Applications in Complex Natural Structures

The drive for robust skeletonization is largely motivated by the need to analyze complex, branching structures in nature and industry:

  • Medical Imaging: Tools like AirQuant, developed by Ashkan Pakzad and colleagues, use skeletons to quantify lung scarring. In these cases, a "false branch" isn't just a geometric error—it's a false positive for a disease.
  • Art and Heritage: Oleksii Sidorov and Jon Yngve Hardeberg treat the cracks in old paintings (craquelure) as a skeletal graph in Craquelure as a Graph, using the resulting topology to identify the origin of artworks.
  • Biological Modeling: In A Novel Technique... to Improve Leaf Detection in Maize, Nazifa Azam Khan shows how pruning "spurs" is essential for tracking plant growth, where noise could otherwise be misinterpreted as a new leaf.

Toward a Multi-Scale Theory of Skeletons

Recent work has begun to unify these robust techniques into a formal "scale-space" theory. Julia Gierke proposes in Skeletonisation Scale-Spaces that we should not view a skeleton as a single static object, but as a hierarchy of representations. By systematically removing or adding skeletal points based on their significance, we can create a "scale-space" of skeletons that are inherently robust to noise and can be tuned to the specific level of detail required for a task.

By combining these theoretical guarantees, such as those provided by Vitaliy Kurlin for unorganised point clouds in Skeletonisation Algorithms for Unorganised Point Clouds with Theoretical Guarantees, with modern optimization and learning, the field is finally turning the medial axis from a fragile mathematical curiosity into a reliable tool for structural analysis.

Papers in this version

Showing 60 of 99 papers.

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 1, frozen text):

@misc{lacuna_robust_skeletonization_for_complex_geome_8d0bfceb_v1,
  title = {Robust Skeletonization for Complex Geometric Structures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-skeletonization-for-complex-geometric-structures/rdr_8d0bfcebaca847acbaeef1a33c9c756b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}