Continuous Skeleton Smoothing for Robust Shape Representation

17 concepts9 papers1 proposal

When we look at a complex shape—whether it is a branching blood vessel, a handwritten character, or a human silhouette—our brains naturally extract a simplified, topological "core." We intuitively understand the shape's structure by its centerline and thickness rather than its exact pixel boundaries.

In computer vision and medical imaging, however, raw data is rarely clean. Boundary noise, pixel discretization, and environmental degradation often turn smooth biological or geometric structures into jagged, fragmented artifacts. Traditional pixel- or voxel-based representations struggle with this noise because they treat every data point independently, failing to capture the global continuity of the underlying object.

To solve this, a powerful paradigm has emerged: reducing shapes to their topological skeletons, smoothing those skeletons using continuous mathematical functions (such as B-splines), and then analyzing or "re-inflating" them. This approach decouples an object's core topology from its surface-level noise, providing a robust, interpretable, and mathematically differentiable foundation for shape analysis.


Decoupling Topology from Surface Noise

Why do we extract a skeleton instead of smoothing the shape's boundary directly? Boundaries are highly sensitive to local noise. A small crack or a piece of dirt on an image can drastically alter a boundary's geometry, but it rarely changes the overall topology of the object.

By compressing a shape down to its medial axis or centerline, we isolate its structural essence. As demonstrated by Fausto Giunchiglia et al. in the context of highly degraded ancient oracle bone inscriptions, using a simplified structural skeleton as an intermediate representation allows generative models to reconstruct clean versions of complex symbols without being misled by surface artifacts like cracks or stains.

[Noisy Shape] ──> [Extract Medial Axis] ──> [Prune & Smooth Skeleton] ──> [Re-inflate / Analyze]

Once a skeleton is extracted, it can be cleaned by pruning away small, spurious branches caused by boundary noise. Cristian Sminchisescu et al. leveraged this mechanism for human pose estimation from noisy silhouettes. By extracting the medial axis, pruning minor noise-induced branches, and "re-inflating" the remaining skeleton, the authors created an idealized, denoised version of the target shape that preserved the core human structure while discarding high-frequency artifacts.


From Discrete Points to Continuous Splines

Even after pruning, a raw skeleton extracted from a pixel grid remains discrete and jagged. If you try to calculate geometric properties—like the curvature of a spine or the bending of a blood vessel—on a discrete set of points, the results will be highly unstable due to sampling artifacts.

To bridge this gap, researchers fit the discrete skeleton points to continuous parametric curves, most notably B-splines.

Discrete Points:   o ── o ── o ── o ── o  (Jagged, non-differentiable)
                               ↓
B-Spline Curve:    ~~~~~~~~~~~~~~~~~~~~~  (Smooth, continuous derivatives)

B-splines offer two critical advantages for shape representation:

  1. Inherent Smoothness and Flexibility: They enforce continuity across the entire curve, making them highly resilient to localized errors or missing data.
  2. Analytical Derivatives: Because a B-spline is a continuous function, you can mathematically calculate its exact derivatives at any point, yielding instantaneous geometric metrics without relying on crude finite differences.

This mathematical elegance makes splines highly effective across diverse domains:

  • Anatomical Modeling: Hao Wang et al. represented the human spine as a B-spline curve by first predicting initial centerline points and then estimating spline parameters. This approach inherently captures the spine's natural flexibility and smoothness, making curvature estimation robust even in low-quality X-ray images.
  • Vascular Analysis: In cerebrovascular modeling, Wang Xing-Ce et al. fitted B-splines to extracted vessel skeletons. This continuous representation allowed for precise resampling along the vessel's central axis, eliminating errors introduced by discrete voxel sampling.
  • Scene Text Detection: For highly curved, ribbon-like text lines, Chuang Yang 0003 et al. used rotating coordinate systems and polynomial fitting to ensure structural continuity during centerline extraction, preventing jagged artifacts in predicted text skeletons.

The Power of Continuous Geometric Metrics

Once a shape is represented as a smooth, continuous skeleton, we can extract richer geometric insights than traditional discrete metrics allow.

Consider the problem of analyzing neuronal structures. Traditional metrics like tortuosity (the ratio of the curve length to the distance between its endpoints) require defining arbitrary start and end points, reducing an entire intricate path to a single discrete number.

By contrast, Thomas L. Athey et al. fit B-splines to 3D axonal arbors to calculate continuous, point-wise metrics like curvature and torsion. Because these metrics are defined at every single point along the continuous curve, they provide a complete, localized characterization of the axon's shape, revealing distinct geometric differences between primary and terminal branches that discrete metrics miss.

Similarly, Xiaozheng Zhang et al. introduced the Width Distribution descriptor for recognizing articulated objects. By fitting the medial axis to B-splines, the authors ensured that calculations of local thickness (the "flesh" around the bone) were based on a stable, continuous coordinate system, making object recognition invariant to boundary discretization.


Re-Inflation: Putting the Flesh Back on the Bones

A skeleton alone is just a line. To reconstruct the full 3D volume or surface mesh of an object, we must "re-inflate" the smooth skeleton by combining it with local thickness or geometric features.

[Smooth Centerline] + [Continuous Radius Function] ──> [Smooth 3D Mesh]

This concept is beautifully realized in vascular vectorization. Xiaoyu Yang et al. segmented coronary arteries by identifying centerlines and sampling cross-sections at regular intervals. By applying a smoothing filter directly to the radii of these cross-sections along the continuous centerline, their model generates anatomically continuous 3D meshes, entirely eliminating the "broken vessel" problem that plagues traditional voxel-based segmentation.

On a theoretical level, this synthesis of topology and geometry is formalized by frameworks like Decorated Reeb Spaces. As explored by Justin Curry et al., these hybrid descriptors combine the structural "skeleton" of a Reeb space with the local geometric "flesh" of persistent homology. This mathematical framework proves that simplifying complex data via topological skeletons can be achieved stably, clarifying high-level organizational structures without sacrificing critical local geometric information.


Conclusion

Continuous skeleton smoothing works because it respects the underlying physics and biology of the objects it models. Blood vessels, spines, and written strokes do not consist of independent, jagged pixels; they are continuous, smooth trajectories with varying thickness. By mapping discrete data onto parametric curves like B-splines, this research direction provides a mathematically rigorous way to filter out noise, compute exact geometric properties, and reconstruct structurally sound shapes.

Key Foundational Work

Human Pose Estimation from Silhouettes - A Consistent Approach Using Distance Level Sets (2002) — contributed 2 concept elements to this direction.
Width Distributions for Shape Description (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_smooth_curve_fitting_c4902d67,
  title = {smooth curve fitting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/smooth-curve-fitting/rdr_c4902d671eb64a03a1f07ae3e92684f1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_smooth_curve_fitting_c4902d67_v1,
  title = {smooth curve fitting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/smooth-curve-fitting/rdr_c4902d671eb64a03a1f07ae3e92684f1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake