A Second-order Total Variation Metric on the Space of Immersed Curves

Giacomo Nardi, Gabriel Peyré, François‐Xavier Vialard

Abstract

This paper studies the space of $BV^2$ planar curves endowed with the $BV^2$ Finsler metric over its tangent space of displacement vector fields. Such a space is of interest for applications in image processing and computer vision because it enables piecewise regular curves that undergo piecewise regular deformations, such as articulations. The main contribution of this paper is the proof of the existence of a shortest path between any two $BV^2$ curves for this Finsler metric. The method of proof relies on the construction of a martingale on a space satisfying the Radon-Nikodym property and on the invariance under reparametrization of the Finsler metric. This method applies more generally to similar cases such as the space of curves with $H^k$ metrics for $k\geq 2$ integer. When $k \geq 2$ is integer, this space has a strong Riemannian structure and is geodesically complete. Thus, our result shows that the exponential map is surjective, which is complementary to geodesic completeness in infinite dimensions. We propose a finite element discretization of the minimal geodesic problem, and use a gradient descent method to compute a stationary point of a regularized energy. Numerical illustrations shows the qualitative difference between $BV^2$ and $H^2$ geodesics.

Introduction

In fields like computer vision and medical imaging, the task of "shape registration"—finding an optimal transformation from one shape to another—is fundamental. This problem is typically framed as finding a shortest path, or geodesic, in a "shape space" where each point represents a curve. While traditional approaches often use Sobolev ($H^k$) metrics, these can be overly restrictive, enforcing global smoothness that fails to capture the piecewise regular nature of articulated objects, such as a bending human arm or mechanical joints.

This paper provides a rigorous mathematical proof for the existence of shortest paths between planar curves using a second-order bounded variation ($BV^2$) metric. By moving beyond smooth Sobolev metrics to this Finsler framework, the authors enable shape deformations that can handle articulations and sharp corners—essential for realistic computer vision tasks like tracking human movement or mechanical parts.

The $BV^2$ Shape Space

The space of curves considered is $BV^2$, which allows for "jumps" in the second derivative. This mathematical property is crucial for applications: it means a curve can bend sharply at specific points while remaining rigid elsewhere.

An example of an immersed curve in the plane, representing a point in the $BV^2$ shape space.

Unlike standard $L^2$ metrics, which are degenerate in infinite dimensions (allowing the distance between distinct curves to be zero), the $BV^2$ metric provides a well-defined distance. However, $BV^2$ is a Finsler space, meaning the metric is defined by a norm on the tangent space rather than an inner product. This adds significant complexity to the proof of existence for geodesics.

Proving the Existence of Minimal Geodesics

The core contribution is the proof that a path minimizing the geodesic energy always exists between any two curves in the $BV^2$ space. The authors utilize the direct method of the calculus of variations, which requires showing that a minimizing sequence of paths has a convergent subsequence.

The Martingale Approach

A major technical hurdle is that the $BV^2$ space lacks the Radon-Nikodym Property (RNP), which is usually necessary for standard weak convergence proofs in infinite-dimensional spaces. To overcome this, the authors developed a novel approach using a dyadic decomposition of the path. By constructing a martingale on a related space that does satisfy the RNP, they were able to identify a limit path that serves as the minimal geodesic.

Reparameterization Invariance

The paper distinguishes between "parameterized" curves (where the representation depends on the mapping from an interval) and "geometric" curves (where only the image in the plane matters). The authors prove that the $BV^2$ metric is invariant under re-parameterization, ensuring that the existence results hold for the intrinsic geometry of the shapes.

Numerical Implementation and Results

To demonstrate the practical utility of the $BV^2$ metric, the authors proposed a finite element discretization using piecewise affine curves. Because the $BV^2$ norm is non-smooth and non-differentiable, they introduced a regularized energy functional, smoothed by a parameter $\epsilon$, to enable optimization via gradient descent.

Qualitative Comparison

Numerical experiments on synthetic models—such as morphing a square or a "V" shape—highlight the advantages of the $BV^2$ approach. While $H^2$ (Sobolev) geodesics tend to "round off" corners during the transition to maintain global smoothness, $BV^2$ geodesics preserve the integrity of sharp features. This "piecewise rigidity" is highly desirable for tracking articulated systems.

Convergence and Surjectivity

The authors proved that as the smoothing parameter $\epsilon$ approaches zero, the numerical solutions converge to the true minimal geodesic. Furthermore, for the smoother $H^k$ cases, they established that the exponential map is surjective, meaning any curve in the space can be reached from any other by following a geodesic—a significant theoretical result for infinite-dimensional manifolds.

Limitations and Future Directions

Despite the theoretical success, several challenges remain:

  • Local Minima: The energy functional is highly non-convex, and gradient descent can easily become trapped in local minima. Careful initialization is required.
  • Computational Cost: The non-smooth nature of the Finsler metric makes it more computationally demanding than Riemannian alternatives.
  • Euler-Lagrange Equations: Deriving generalized Euler-Lagrange equations for the $BV^2$ case is difficult due to the lack of differentiability.

Future work may focus on designing Finsler metrics that even more strongly favor piecewise rigid motions, potentially improving the tracking of complex mechanical and biological systems.

Summary updated

Abstract

This paper studies the space of $BV^2$ planar curves endowed with the $BV^2$ Finsler metric over its tangent space of displacement vector fields. Such a space is of interest for applications in image processing and computer vision because it enables piecewise regular curves that undergo piecewise regular deformations, such as articulations. The main contribution of this paper is the proof of the existence of a shortest path between any two $BV^2$ curves for this Finsler metric. % The method of proof relies on the construction of a martingale on a space satisfying the Radon-Nikodym property and on the invariance under reparametrization of the Finsler metric. This method applies more generally to similar cases such as the space of curves with $H^k$ metrics for $k\geq 2$ integer. When $k \geq 2$ is integer, this space has a strong Riemannian structure and is geodesically complete. Thus, our result shows that the exponential map is surjective, which is complementary to geodesic completeness in infinite dimensions. We propose a finite element discretization of the minimal geodesic problem, and use a gradient descent method to compute a stationary point of a regularized energy. Numerical illustrations shows the qualitative difference between $BV^2$ and $H^2$ geodesics.
BibTeX
@misc{nardi2014second,
  author = {Giacomo Nardi and Gabriel Peyré and François‐Xavier Vialard},
  title = {{A Second-order Total Variation Metric on the Space of Immersed Curves}},
  year = {2014},
  eprint = {1402.6504},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1402.6504},
}
Figure