Introduction
In computer graphics, parameterization is the fundamental process of mapping a 2D coordinate system onto a 3D surface. This is essential for tasks ranging from texture mapping and decal placement to drawing curves on complex models. Historically, these methods have been "representation-specific," meaning a technique optimized for clean triangle meshes would often fail when applied to noisy point clouds or modern neural implicit functions.
Researchers have recently introduced a robust, representation-agnostic algorithm designed to create high-quality local 2D maps regardless of the underlying geometry. By shifting the focus from global optimization to local, output-sensitive processing, this method enables interactive design on massive datasets and unconventional surface types.
Methodology: Tracing and Smoothing
The core of the approach is a two-stage pipeline rooted in the concept of the Exponential Map, which maps vectors in a tangent plane to paths on a surface.
1. Radial Tracing with Substepping
Starting from a chosen origin point, the algorithm traces several "geodesic-like" paths outward. To maintain accuracy on surfaces with sharp corners or high curvature, the researchers implemented a curvature-sensitive substepping procedure. This ensures that the paths remain strictly on the surface and maintain consistent lengths, preventing the "foldovers" common in simpler projection methods.
2. Holonomy Smoothing
The "secret sauce" of this method is holonomy smoothing. On curved surfaces, paths that start parallel will naturally diverge or converge (a phenomenon described by the Jacobi equation). This leads to angular distortion. The authors solve a 1D Dirichlet problem to "nudge" the rotation of these radial paths, effectively smoothing out the distortion so the paths behave as if they were on a surface with constant curvature.
3. Spline Interpolation
Finally, the discrete points from the traced paths are fused into a continuous surface using periodic and natural cubic splines. This requires only a minimal interface with the geometry: the ability to query a Signed Distance Function (SDF) and project points onto the surface.
Results and Performance
The method was evaluated against a diverse benchmark of over 100 surfaces, including biological meshes (like MRI-scanned brains), massive LiDAR-scanned terrains with over 29 million points, and neural implicit models (NGLOD architecture).
Key findings include:
- Lower Distortion: The method consistently outperformed the "Vector Heat Method" and "Discrete Exponential Maps" in Symmetric Dirichlet and LSCM metrics, particularly on poorly sampled surfaces.
- Efficiency: Most local maps were generated in under 1.5 seconds on consumer-grade hardware (M1 MacBook Pro), making it suitable for interactive tools.
- Scalability: Because the algorithm is output-sensitive, it only processes the local area where a decal is placed, allowing it to handle massive models that would crash global parameterization tools.
Conclusion and Future Work
While currently restricted to topological discs (it cannot yet "wrap around" complex handles or bridge large holes), the method represents a significant step toward representation-independent geometry processing. Future work may involve an "atlas" approach to stitch multiple local maps together, potentially enabling complex physical simulations directly on neural implicit surfaces.
Summary updated