The Method of Moving Frames for Surface Global Parametrization

Guillaume Coiffier, Etienne CormanACM Trans. Graphics 2023 · DOI

Introduction

Mapping complex 3D surfaces to 2D planes—global parametrization—is a cornerstone of computer graphics, essential for tasks like texture mapping and quad-meshing. However, flattening a non-developable surface inevitably introduces distortion or requires "singularities" (points like the poles of a globe where the map is not smooth). Traditionally, this has been a two-step "chicken and egg" problem: one must first decide where to place singularities and cuts (topology), and then solve for the mapping (geometry). Poor initial placement of singularities leads to high geometric distortion, yet optimizing both simultaneously often requires computationally expensive integer programming.

Presented at SIGGRAPH 2023, this work introduces a unified framework that treats singularity placement and geometric mapping as a single continuous optimization problem. By leveraging Cartan’s Method of Moving Frames, the authors eliminate the need for pre-defined cuts or discrete integer variables, achieving state-of-the-art distortion minimization.

Methodology: Cartan’s Moving Frames

The core innovation lies in the use of local frames rather than global coordinates. The method employs Cartan’s "first structure equation" as a local integrability condition, describing how local coordinates evolve as a frame moves across the surface.

To accommodate the specific requirements of quad-meshing, the theory is extended to symmetric frame fields ($n$-RoSy). This allows the optimization to "quantize" singularities—for instance, forcing them to be multiples of 90 degrees—within a continuous framework, bypassing the NP-hard complexity of integer programming. Furthermore, the authors utilize vertex-based charts, subdividing mesh triangles into quadrilaterals. This discretization allows singularities to emerge anywhere on the surface, making the method robust even on coarse or poorly sampled meshes. The final formulation is a non-linear least squares problem that minimizes distortion energies (such as As-Rigid-As-Possible) while strictly satisfying the structure equations.

Key Results

The algorithm was evaluated on a diverse database of 274 triangular meshes, including 131 smooth models and 143 complex CAD models from the Mambo dataset. It achieved a high success rate of 93%.

Comparison of parametrization results using different optimization objectives, demonstrating the trade-offs between cone count and geometric distortion.

As shown in the comparison above, the method allows singularities (cones) to be "discovered" naturally by the optimization to minimize specific distortion metrics. Compared to traditional two-step baselines, this approach consistently produces lower "stretch" distortion. On a benchmark of 62 meshes, the method generated 15% fewer cones on average while maintaining superior mapping quality. The framework is highly versatile, supporting various constraints such as alignment with sharp feature edges or specific distortion metrics like area-preserving or isometric mappings.

Discussion and Future Work

A primary advantage of this unified approach is that the link between cone placement and distortion is never broken, leading to significantly improved results. By using a "power vector" representation for frame fields, valid singularity indices are enforced through continuous optimization.

However, some limitations remain. While the method places cones optimally, it does not guarantee they will lie on an integer grid, necessitating a post-processing step for perfect quad-meshing. Additionally, the increased number of variables per vertex makes it slower than simpler, less accurate methods. Future work may involve extending this theory to 3D tetrahedral meshes for hexahedral remeshing, as the underlying moving frames theory is not restricted to two dimensions.

Summary updated

BibTeX
@article{coiffier2023method,
  author = {Guillaume Coiffier and Etienne Corman},
  title = {{The Method of Moving Frames for Surface Global Parametrization}},
  journal = {ACM Trans. Graphics},
  year = {2023},
  doi = {10.1145/3604282},
}

Loading related content...

Figure