Feature Extraction on Meshes using CSRBF Fitting

Yutaka Ohtake, Alexander Belyaev, Hans‐Peter SeidelKeisan 2004 · DOI

Abstract

This paper presents a robust methodology for extracting ridge and valley lines from 3D triangle meshes. By fitting a smooth, implicit surface to discrete mesh data using Compactly Supported Radial Basis Functions (CSRBF), the approach enables the analytical calculation of high-order derivatives. This transition from discrete geometry to a continuous mathematical representation results in cleaner, more stable feature detection, even when processing complex or noisy geometric shapes.

Introduction

In computer graphics and geometric modeling, identifying "features" such as ridges (mountain-like crests) and valleys (trough-like lines) is essential for shape recognition, reverse engineering, and mesh simplification. However, detecting these lines is mathematically demanding, requiring the calculation of first and second derivatives of principal curvatures.

On a standard triangle mesh—a collection of flat faces—these derivatives are extremely sensitive to noise and discretization errors. Calculating them directly from the mesh is analogous to finding the exact peak of a staircase; the discrete steps make the mathematics unstable and unreliable. This paper addresses these challenges by shifting the problem from the discrete domain to a continuous one.

Methodology

The proposed pipeline consists of five primary stages designed to stabilize feature extraction:

  1. Implicit Surface Fitting (CSRBF): Instead of working directly on triangles, the method fits a smooth implicit surface ($f(x) = 0$) to the mesh vertices and normals. Using CSRBFs combined with a "Partition of Unity" approach allows for efficient representation of complex, large-scale geometry.
  2. Surface Projection: Since original mesh vertices may not lie perfectly on the new mathematical surface, Newton’s method is employed to project vertices onto the $f(x) = 0$ surface.
  3. Analytical Derivative Calculation: With the surface defined by a mathematical function, standard calculus is used to analytically derive principal curvatures ($k_{max}, k_{min}$) and their derivatives along principal directions.
  4. Feature Extraction: Ridge and valley points are identified where the derivative of curvature along its principal direction is zero, subject to specific threshold conditions.
  5. Filtering and Cleaning: To mitigate "noise" features caused by minor surface fluctuations, the algorithm calculates the integral of the principal curvature along each detected line segment. Segments with low total "strength" are discarded.
Figure 4: A multi-stage visualization of the feature extraction and cleaning process on a geometric test object, showing the progression from raw surface to refined feature lines.

Key Insights and Results

The core innovation of this research is the "Smooth Skin" approach. By treating the mesh as a proxy for an underlying smooth shape and reconstructing that surface, researchers gain access to continuous calculus tools that are far more stable than discrete approximations.

Performance and Stability

The method successfully extracts smooth, continuous ridge and valley lines that align with human visual perception. Using CSRBF with Partition of Unity is significantly faster than alternatives like Moving Least Squares (MLS), as it avoids redundant optimizations for every evaluation point.

Application to Complex Geometry

The algorithm demonstrates high performance on non-trivial, real-world geometry. While simple primitives validate the mathematical correctness, the method's true value is seen in its ability to handle high-resolution, organic meshes where discrete curvature estimation typically fails.

Figure 6: Application of the feature extraction method to a complex dental mesh, demonstrating its effectiveness on organic, real-world geometry.

Limitations and Future Work

Despite its robustness, the initial detection phase remains sensitive to small surface fluctuations, which can lead to fragmented feature sets before the cleaning stage. Additionally, while faster than some implicit methods, fitting a global or semi-global surface to meshes containing millions of triangles still carries a higher computational cost than local mesh operators. Future work may focus on optimizing the initial detection to reduce fragmentation and further streamlining the fitting process for massive datasets.

Summary updated

BibTeX
@inproceedings{ohtake2004feature,
  author = {Yutaka Ohtake and Alexander Belyaev and Hans‐Peter Seidel},
  title = {{Feature Extraction on Meshes using CSRBF Fitting}},
  booktitle = {Keisan},
  year = {2004},
  doi = {10.1299/jsmecmd.2004.17.211},
}
Figure