The Vector Space of Convex Curves: How to Mix Shapes

Dongsung HuharXiv 2015 · arXiv

Abstract

This paper introduces a "log-radius profile" representation for convex curves, allowing shapes to be treated as elements in a formal vector space. By representing curves through the logarithm of their radius of curvature, we can "mix" shapes via addition and scalar multiplication in a mathematically elegant and visually intuitive way, overcoming the distortion issues found in traditional Fourier Descriptor methods.

Introduction: The Challenge of Shape Representation

In computer vision and human perception research, representing and combining shapes is a fundamental challenge. The standard approach uses angle profiles (often analyzed via Fourier Descriptors), which describe how the orientation of a curve changes along its arc length. However, angle profiles have a major flaw: they don't "mix" well. Combining the features of two shapes by adding their angle profiles often results in distorted curves that lose the characteristic features of the originals. This limitation makes it difficult to perform meaningful shape synthesis or to model how the brain might interpolate between different visual objects.

The Log-Radius Profile

The author proposes a dual representation called the log-radius profile. For convex curves—those that always turn in one direction—every point can be uniquely identified by its orientation angle $\theta$. The derivative of the curve with respect to this angle gives the local radius of curvature, $r(\theta)$. The shape is then characterized by the function:

$$l(\theta) \equiv \log r(\theta)$$

This framework turns the set of convex curves into a Hilbert space, which is mathematically isomorphic to the space of square-integrable functions ($L_2$).

Frequency as Symmetry

A key insight of this representation is the relationship between frequency and rotational symmetry. By defining elementary shapes using sinusoidal log-radius profiles ($l(\theta) = \epsilon \sin(\nu\theta)$), the frequency $\nu$ directly corresponds to the shape's symmetry.

Elementary shapes generated via sinusoidal log-radius profiles, where the frequency $\nu$ corresponds to rotational symmetry.

As shown in the figure above, $\nu=2$ creates an ellipse, while higher integers create rounded polygons such as triangles ($\nu=3$) and squares ($\nu=4$). Fractional values of $\nu$ produce more complex, self-intersecting or spiral-like forms.

Vector Space Operations: Mixing Shapes

Because these profiles are functions in a vector space, we can define formal operations for addition and scalar multiplication:

  • Addition: $\Gamma_1 + \Gamma_2 \longleftrightarrow l_1(\theta) + l_2(\theta)$. This acts as a "generalized scaling" where one shape modifies the local scale of another.
  • Scalar Multiplication: $a \cdot \Gamma \longleftrightarrow a \cdot l(\theta)$. This amplifies or dampens the "expressiveness" of a shape's features (e.g., making a triangle more or less "pointy") without destroying its underlying symmetry.

Unlike traditional methods, adding profiles in this space preserves the features of both parent shapes in a way that looks "correct" to a human observer.

Addition of a spiral profile to a pentagonal profile results in a "pentagonal spiral," demonstrating intuitive feature preservation.

Key Results and Comparison

The paper successfully formalizes a vector space for convex curves, providing a rigorous foundation for shape mixing. Qualitative results demonstrate that the log-radius approach remains stable and meaningful even as complexity increases. In contrast, traditional Fourier Descriptors often fail to represent unique, consistent shape features as coefficients increase. Furthermore, the angle coordinate $\theta$ is naturally scale-invariant, ensuring the representation remains consistent regardless of the object's overall size.

Limitations and Future Work

The current model is strictly limited to convex curves. Representing non-convex or more complex shapes would require extending this logic. Future work may involve generalizing this method to higher-dimensional convex surfaces using normal vectors and the inverse Gauss map. Additionally, the author suggests this representation might mirror biological processes, such as how the brain perceives motion or generates curved hand movements, providing a potential bridge between geometry and neuroscience.

Summary updated

Abstract

We present a novel, log-radius profile representation for convex curves and define a new operation for combining the shape features of curves. Unlike the standard, angle profile-based methods, this operation accurately combines the shape features in a visually intuitive manner. This method have implications in shape analysis as well as in investigating how the brain perceives and generates curved shapes and motions.
BibTeX
@misc{huh2015vector,
  author = {Dongsung Huh},
  title = {{The Vector Space of Convex Curves: How to Mix Shapes}},
  year = {2015},
  eprint = {1506.07515},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1506.07515},
}
Figure