Abstract
In shape analysis and computer vision, measuring the distance between curves is fundamental for tasks like morphing and segmentation. While the $H^0$ (or $L^2$) metric is the most intuitive choice, it suffers from a "vanishing distance" paradox where the distance between any two distinct curves is zero. This paper explores the geometry of the manifold of curves and proposes a class of "conformal" metrics. These metrics preserve the computational efficiency of second-order methods while restoring a meaningful distance structure, providing a robust framework for Level Set implementations.
Introduction
The mathematical representation of shapes often involves treating the set of all possible curves as a high-dimensional manifold $M$. For curves parameterized as $c: S^1 \to \mathbb{R}^n$, the tangent space $T_cM$ at a curve $c$ consists of all possible infinitesimal deformations $h$. To perform operations like shape morphing or finding the "average" of multiple shapes, we must define a Riemannian metric $F(c, h)$ to calculate the energy of these deformations and identify the shortest paths (geodesics) between shapes.
The Vanishing Distance Paradox
The most straightforward metric is the $H^0$ metric, defined as the integral of the squared velocity of every point on the curve: $F(c, h) = \int |h|^2 ds$. However, as first noted by Mumford, this metric is mathematically "broken." In the space of curves, one can introduce high-frequency "wiggles" that allow a curve to deform into another with arbitrarily low energy. Consequently, the geodesic distance between any two distinct shapes under the pure $H^0$ metric is effectively zero.
Proposed Solutions: Constraints and Conformal Metrics
The authors investigate two primary methods to resolve this pathology:
- Curvature Constraints: By imposing a strict limit on the total curvature of the curves, the $H^0$ metric can be "saved," yielding positive distances. This approach highlights that the paradox arises from unconstrained parameterization and excessive bending.
- Conformal Metrics: The paper's core contribution is the introduction of conformal modifications to the $H^0$ metric. By multiplying the standard metric by a "conformal factor" $\phi(c)$—typically a function of the curve's total length $L$—the space is "stretched" in a way that penalizes the problematic high-frequency deformations. A common choice is $\phi(c) = e^{\lambda L}$, which effectively prevents the zero-distance collapse.
Implementation and Level Set Methods
A significant advantage of the conformal approach is its compatibility with Level Set Methods. Previous attempts to fix the $H^0$ metric relied on "higher-order" metrics (like the Michor-Mumford metric) involving fourth-order derivatives. While theoretically sound, these are computationally expensive and numerically unstable.
In contrast, the proposed conformal metrics remain second-order. The authors derive a minimizing flow that can be embedded into a 4D scalar function, allowing for the use of standard, stable numerical solvers. This makes the conformal metric a practical "drop-in" improvement for existing computer vision pipelines, such as those used in medical imaging or object tracking.
Conclusion
By addressing the fundamental geometric flaws of the $H^0$ metric, this work provides a robust mathematical foundation for shape analysis. The conformal metric framework offers a rare balance between theoretical rigor—solving the vanishing distance paradox—and practical utility, enabling smooth and physically intuitive shape morphing without the computational burden of higher-order methods.
Full Citation: A. Yezzi and A. C. G. Mennucci, "Metrics in the space of curves," arXiv:math/0412454v2 [math.DG], 2005.
Summary updated