Curve cuspless reconstruction via sub-Riemannian geometry

Ugo Boscain, Remco Duits, Francesco Rossi, Yuri L. Sachkov

Abstract

We consider the problem of minimizing $\int_{0}^L \sqrt{\xi^2 +K^2(s)}\, ds$ for a planar curve having fixed initial and final positions and directions. The total length $L$ is free. Here $s$ is the variable of arclength parametrization, $K(s)$ is the curvature of the curve and $\xi > 0$ a parameter. This problem comes from a model of geometry of vision due to Petitot, Citti and Sarti. We study the existence of local and global minimizers for this problem. We prove that if for a certain choice of boundary conditions there is no global minimizer, then there is neither a local minimizer nor a geodesic. We finally give properties of the set of boundary conditions for which there exists a solution to the problem.

Introduction

When the human visual system encounters interrupted or obscured lines, the primary visual cortex (V1) performs a "filling-in" process. This biological mechanism can be modeled as a variational problem where the brain seeks to minimize an energy functional to find the most plausible path. A famous model by Petitot, Citti, and Sarti frames this as a problem in sub-Riemannian geometry on the manifold $R^2 \times S^1$.

Biological architecture of the visual cortex V1 showing orientation columns and hypercolumns.

A significant challenge in these models is the emergence of "cusps"—points where a curve abruptly reverses direction. In human perception, we perceive smooth, continuous paths rather than sharp reversals. This paper addresses the (P_curve) problem: finding the minimum-energy path while strictly forbidding these cusps by enforcing a constant velocity constraint.

Methods

The authors employ Optimal Control Theory and Sub-Riemannian Geometry to analyze the problem. They seek to minimize the functional $\int \sqrt{\xi^2 + K^2(s)} ds$, where $K$ is curvature and $\xi$ is a weighting parameter. By enforcing a constant velocity ($u=1$), they prevent the direction reversals that cause cusps.

Because the optimal controls (curvature) might not be bounded (non-Lipschitz), the study utilizes a generalized version of the Pontryagin Maximum Principle (PMP). To solve the constrained problem, the authors first analyze a related mechanical problem (P_MEC) where cusps are allowed, using its symmetry and optimality properties to determine if solutions can be re-parametrized into "cuspless" forms.

Results

The most striking finding is the All-or-Nothing Theorem: for any given set of boundary conditions, either a global minimizer exists, or there are no local minimizers or geodesics at all. This establishes a binary existence condition for smooth curve reconstruction.

Numerical mapping of the existence set S for boundary conditions.

The authors numerically mapped the "existence set" $S$, which characterizes the boundary conditions for which a solution exists. They proved that $S$ is arc-connected and non-compact. This mapping reveals "forbidden zones" where the geometry of the starting and ending points is too extreme for a smooth reconstruction, providing a mathematical explanation for the limits of human visual integration.

Conclusion

This work bridges abstract sub-Riemannian geometry and human visual perception by providing a rigorous proof for when smooth curve reconstructions are possible. While the model treats the visual cortex as a continuous space, it offers a robust framework for understanding how V1 processes geometric information. Future work will focus on analytical descriptions of the existence set boundaries and extending the model to higher dimensions or non-Euclidean surfaces.

Summary updated

Abstract

We consider the problem of minimizing $\int_{0}^L \sqrt{ξ^2 +K^2(s)}\, ds $ for a planar curve having fixed initial and final positions and directions. The total length $L$ is free. Here $s$ is the variable of arclength parametrization, $K(s)$ is the curvature of the curve and $ξ>0$ a parameter. This problem comes from a model of geometry of vision due to Petitot, Citti and Sarti. We study existence of local and global minimizers for this problem. We prove that if for a certain choice of boundary conditions there is no global minimizer, then there is neither a local minimizer nor a geodesic. We finally give properties of the set of boundary conditions for which there exists a solution to the problem.
BibTeX
@inproceedings{boscain2014curve,
  author = {Ugo Boscain and Remco Duits and Francesco Rossi and Yuri L. Sachkov},
  title = {{Curve cuspless reconstruction via sub-Riemannian geometry}},
  booktitle = {ESAIM},
  year = {2014},
  doi = {10.1051/cocv/2013082},
}

Loading related content...

Figure