Natural Boundary Conditions for Smoothing in Geometry Processing

Oded Stein, Eitan Grinspun, Max Wardetzky, Alec JacobsonACM Trans. Graphics 2018 · DOI

TL;DR

This paper introduces a new smoothness energy for geometry processing, based on the squared Hessian, that naturally handles boundaries without introducing bias. This "squared Hessian energy" offers more desirable behavior than traditional methods, especially for irregular domains, leading to more natural smoothing and interpolation results.

What Problem Does This Paper Address?

In geometry processing, smoothing is crucial for tasks like denoising, interpolation, and shape optimization. A common approach uses smoothness energies, often based on the Laplacian operator. However, when dealing with domains that have boundaries where values aren't explicitly known (free boundaries), standard implementations of these energies implicitly enforce "low-order" boundary conditions. These conditions can introduce undesirable biases, causing the smoothed results to unnaturally follow the boundary's shape, as illustrated in Figure 1. The authors aim to develop a smoothness energy whose natural boundary conditions avoid this bias, allowing the data itself to dictate the behavior near the boundary.

Methods

The core contribution is the proposal and analysis of the squared Hessian energy ($E_{H^2}$) as a smoothness energy. Unlike the squared Laplacian energy ($E_{\Delta^2}$), $E_{H^2}$ has natural boundary conditions that are "high-order" and do not impose a strong prior on the function's behavior near the boundary.

The authors derive these natural boundary conditions using calculus of variations and Green's identities. They show that for a parameterized family of energies that includes both $E_{\Delta^2}$ and $E_{H^2}$, the natural boundary conditions depend on the parameter $\alpha$. For $E_{H^2}$ ($\alpha=1$), these conditions lead to a null space consisting of linear functions, which is intuitive for tasks where boundaries should be ignored.

They then present discretizations of $E_{H^2}$ with its natural boundary conditions using:

  • Finite differences on 2D grids.
  • Mixed finite elements on triangle meshes (both 2D and extended to 3D).

They also explore the relationship between their proposed energy and existing discrete energies, particularly those derived from discrete exterior calculus (DEC) and non-conforming finite elements, highlighting differences in their null spaces and convergence properties.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis and demonstrates its methods on synthetic examples and a range scan of a Nokia cellphone.

Key Insights and Ideas

  • Natural boundary conditions are key: The paper emphasizes that "natural" boundary conditions, derived from the energy minimization itself, are crucial for avoiding unwanted boundary bias, especially when boundaries are not explicitly defined or are meant to be ignored.
  • Hessian vs. Laplacian: The squared Hessian energy ($E_{H^2}$) offers a fundamentally different behavior at boundaries compared to the squared Laplacian energy ($E_{\Delta^2}$). While $E_{\Delta^2}$'s natural boundary conditions can be problematic, $E_{H^2}$'s natural conditions lead to more desirable outcomes.
  • Linear null space for $E_{H^2}$: The null space of $E_{H^2}$ consists of all linear functions. This is a powerful insight for applications like linear subspace design for deformation, as it guarantees that linear transformations are preserved without adding artificial constraints.
  • Improved interpolation and smoothing: The paper demonstrates that using $E_{H^2}$ with its natural boundary conditions leads to smoother interpolations and less biased smoothing, particularly on irregular domains or those with many free boundaries.
  • Discretization matters: The authors show that while some existing discrete energies might seem related, their null spaces and convergence properties can differ significantly, impacting their suitability for certain tasks.

Key Results

  • Theoretical derivation: The paper rigorously derives the natural boundary conditions for the squared Hessian energy.
  • Qualitative improvements: Figures throughout the paper visually demonstrate that $E_{H^2}$ with natural boundary conditions produces smoother results with less boundary bias compared to methods using low-order boundary conditions (e.g., Figure 1, 3, 8, 11, 13).
  • Convergence: The finite difference discretization of $E_{H^2}$ is shown to converge to the analytic solution (Figure 10).
  • Application examples: The paper showcases the effectiveness of $E_{H^2}$ for tasks like data interpolation, smoothing, and linear subspace design for deformation, showing it can handle complex geometries and noisy data more gracefully.

Limitations and Future Work

  • Curvature on surfaces: The current 3D extension of the squared Hessian energy discretization treats surfaces as locally flat. Accounting for curvature (using covariant derivatives) on curved surfaces is identified as a significant area for future work (Section 5.3.1).
  • Boundary-only discretizations: Exploring boundary-only discretizations for flat domains is suggested.
  • Solid objects: Extending the formulation to solid objects and investigating its behavior with point constraints for smooth minimizers is mentioned as future work.
  • Higher-order tensors: The possibility of considering the squared Frobenius norm of third derivatives for even higher-order smoothing is hinted at.
  • Discrete energy convergence: The paper notes that the convergence of some existing discrete energies (like the one from Bergou et al.) to continuous energies with natural boundary conditions is unclear.

Summary updated

BibTeX
@article{stein2018natural,
  author = {Oded Stein and Eitan Grinspun and Max Wardetzky and Alec Jacobson},
  title = {{Natural Boundary Conditions for Smoothing in Geometry Processing}},
  journal = {ACM Trans. Graphics},
  year = {2018},
  doi = {10.1145/3186564},
}
Figure