Curvature Extrema for Shape Representation and Feature Extraction

63 concepts60 papers

When humans look at an object, our eyes are naturally drawn to its corners, turns, and protrusions. In computer vision and geometry processing, this intuition translates to a powerful principle: the most informative parts of a shape or contour are located where its geometry changes most significantly. As argued by Baker in Three-Dimensional Modeling, focusing on points of high curvature or geometric irregularities provides a representation that is far more discriminative and noise-resistant than processing entire continuous contours.

By treating these high-curvature points as discrete "nodes" or informational primitives, algorithms can compress complex shapes, segment boundaries cleanly, and match features across different views. This article explores how curvature-based feature extraction works, moving from simple 2D contours to complex 3D surfaces.

Measuring Curvature in 2D: From Pixels to Vertices

How do we mathematically pinpoint where a curve "bends" when dealing with discrete pixels? Early foundational work by Zahn in A Formal Description for Two-Dimensional Patterns extracted significant vertices along 1D contour lines using a local sliding window and Boolean logic to detect changes in direction.

However, computing derivatives directly on pixel grids is notoriously sensitive to noise. To circumvent this, Fischler et al. proposed a geometric alternative: sliding a chord of a fixed length along a path and measuring the maximum distance between the curve and the chord. Points with high excursions are classified as critical vertices, providing a robust, derivative-free measure of curvature.

      Chord Sliding Mechanism

        /---\  <- High Excursion (Corner Node)
       /     \
  ----*-------*----  <- Sliding Chord

Alternatively, analytical approaches define "cornerness" by looking at intensity gradients. Fang et al. formulated a cornerness measure as the product of the gradient magnitude and the spatial derivative of the gradient direction, ensuring that the metric reaches a local maximum precisely at geometric corners while remaining low along straight edges. To decouple these geometric signatures from distracting background textures, Ravindran et al. extract corners along 1D level lines (iso-intensity curves). This method ensures that the background does not interfere with the feature's signature.

Managing Scale and Noise

Real-world contours are rarely perfectly smooth; pixelation and sensor noise introduce artificial high-frequency jumps. To extract meaningful structural features, algorithms must distinguish true geometric transitions from noise.

One elegant solution is scale-space analysis. Sporring demonstrates in Representing contours as sequences of one dimensional functions that applying Gaussian smoothing allows for the placement of structural knots at semantically important points while effectively filtering out high-frequency pixel-level noise.

Similarly, Singer filters out local noise by selecting curvature extrema using a statistical threshold based on the mean curvature and standard deviation of the specific contour in Significant Feature Detection and Matching in Image Pairs. In pupil tracking, where reflections can corrupt edge data, Santini et al. use a k-cosine chain approximation to identify dominant points, discarding segments that are too straight or fall outside the expected curvature range for a pupil boundary.

Extending to 3D: Ridges, Valleys, and Saddles

The intuition behind 2D curvature extends naturally to 3D surfaces, where curvature is defined by principal curvatures ($k_1$ and $k_2$). Instead of simple corners, 3D feature extraction targets specific geometric topologies like ridges, valleys, and saddle points.

For example, in mesh segmentation, Lee et al. identify natural part boundaries by detecting concave discontinuities where minimum curvature values are highest. These valley regions form thin, skeleton-like graphs that serve as ideal paths for mesh partitioning. Conversely, Hampshire et al. target ridge-like structures—where one principal curvature is high and the other is near zero—to isolate distinct structural landmarks for matching across volumetric colon scans.

3D Feature Type Curvature Condition Common Application
Valley Lines High minimum curvature ($k_{min}$ concavity) Mesh scissoring and part decomposition (Lee et al.)
Ridge Lines One high principal curvature, one near zero Haustral fold matching in CT scans (Hampshire et al.)
Saddle Points Opposing signs ($k_1 \cdot k_2 < 0$) Sinusoidal pattern and soft edge tracking (In the Saddle)

For unorganized point clouds where explicit surface meshes are unavailable, Wang et al. utilize Principal Component Analysis (PCA) to analyze local geometric variation. Points exhibiting high statistical variation in their local neighborhood indicate curvature extrema, allowing the automated identification of complex branching junctions. In intensity space, the authors of In the Saddle: Chasing Fast and Repeatable Features detect interest points by identifying saddle-like intensity profiles (concave in one direction, convex in another), capturing features where traditional blob or corner detectors fail.

Why It Works: The Power of Structural Knots

Curvature-based feature extraction works because it reduces continuous, high-dimensional data into a sparse set of highly descriptive landmarks. This structural compression benefits numerous downstream tasks:

  • Shape Compression and Modeling: Sohel et al. show that segmenting digital shapes based on detected corner points rather than uniform intervals significantly reduces visual distortion. Aligning the endpoints of Bezier curves with these natural features preserves critical geometry while reducing data size by 20-23%. In medical imaging, Yu et al. incorporate mean curvature to automatically identify landmarks on complex bone anatomy, producing highly accurate statistical shape models without relying on manual selection.
  • Object Tracking and Pose Estimation: By focusing on local directional maxima where depth data suggests a prominent protrusion, Breitenstein achieves real-time 3D head pose estimation from single range images. Similarly, Leung et al. extract facial landmarks by tracking where dynamic strips intersect at sharp angles along a profile edge, providing robust tracking across multiple views.
  • Anatomical and Process Mapping: Jiang et al. map local curvature changes along silhouette chain codes directly to human anatomical landmarks (like the shoulder or joints) without relying on global coordinate offsets.

On a theoretical level, Leyton proves in Process Recovery that points of maximum or minimum curvature on a smooth planar curve serve as the terminal points for symmetry axes. This means that curvature extrema inherently encode the physical history and generative processes of a shape. Even in statistics, De Iorio et al. demonstrate that the point of maximum curvature in a probability density function objectively defines the boundary where a distribution's bulk ends and its tail begins.

By anchoring representations to these geometric turning points, vision systems transform raw, unorganized spatial data into structured, semantically meaningful models.

Key Foundational Work

Three-Dimensional Modeling (1977) — contributed 2 concept elements to this direction.
A Formal Description for Two-Dimensional Patterns (1969) — contributed 1 concept elements to this direction.
A Formal Description for Two-Dimensional Patterns (1969) — 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_curvature_based_feature_extraction_a3977297,
  title = {curvature-based feature extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/curvature-based-feature-extraction/rdr_a3977297623a4058a7726dcf929c6d1b},
  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_curvature_based_feature_extraction_a3977297_v1,
  title = {curvature-based feature extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/curvature-based-feature-extraction/rdr_a3977297623a4058a7726dcf929c6d1b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake