In an ideal mathematical world, every 3D shape is represented as a clean, watertight, orientable manifold mesh. In this ideal setting, every edge is shared by exactly two faces, and the local neighborhood around any vertex can be flattened into a 2D disk. This structural regularity makes geometry processing elegant: differential operators are well-behaved, distance metrics propagate smoothly, and defining an "inside" versus an "outside" is trivial.
But real-world data is rarely ideal. Raw LiDAR scans, CAD models, and urban reconstructions are full of "topological defects"—holes, self-intersections, disconnected triangle soups, and non-manifold edges where three or more triangles meet. When traditional geometric algorithms or modern deep learning models encounter these "wild" meshes, the underlying mathematics frequently break down.
To build robust 3D pipelines, researchers are shifting away from strict manifold dependencies. Understanding why this dependency exists—and how modern frameworks bypass it—is central to contemporary geometry processing.
The Comfort of the Manifold Assumption
To understand why non-manifold geometry is so challenging, we must first look at why standard algorithms rely on the manifold assumption in the first place.
Consider the task of shape matching or finding dense correspondences between two deformable objects. Many state-of-the-art frameworks rely on spectral geometry, which translates the shape's physical structure into the frequency domain using the Laplace-Beltrami operator. As Melzi explains in MapTree: Recovering Multiple Solutions in the Space of Maps, calculating these eigenfunctions requires a clean, continuous surface discretization. If an edge connects three faces instead of two, the standard finite-element formulations for the Laplacian lose their mathematical meaning.
Similarly, tasks like computing the thickness of an object via the Shape Diameter Function (SDF) traditionally rely on ray-based intersection logic. As noted by Chen et al. in Fast and robust shape diameter function, if a mesh contains cracks or missing geometry, these rays escape into infinity, causing the thickness metric to fail completely.
Even basic operations like surface partitioning or normal orientation depend heavily on global topological consistency. Xin et al. demonstrate in SurfaceVoronoi: Efficiently Computing Voronoi Diagrams over Mesh Surfaces with Arbitrary Distance Solvers that calculating bisectors and propagating distances for Voronoi diagrams requires predictable face connectivity. When boundaries are open or surfaces are non-orientable, determining a consistent global normal or a clear boundary operator becomes an ill-defined problem (Liu et al. in Fast and Globally Consistent Normal Orientation based on the Winding Number Normal Consistency; Freedman in Surface reconstruction, one triangle at a time).
The Preprocessing Bottleneck
When deep learning entered the 3D domain, it inherited these manifold constraints. Many geometric neural networks require inputs to be closed manifolds to ensure the stability of their convolutional or message-passing operators.
To satisfy this requirement, practitioners traditionally employ a preprocessing step to remesh, repair, or "wrap" raw data into a clean manifold format. However, this creates a severe bottleneck:
- Performance Limits: In Subdivision-based Mesh Convolution Networks, Hu et al. show that forcing arbitrary shapes into regular mesh hierarchies restricts the downstream network's accuracy to the quality of the initial geometric repair.
- Data Scarcity: Generative models, such as the diffeomorphic flow framework proposed by Gupta et al. in Neural Mesh Flow: 3D Manifold Mesh Generation via Diffeomorphic Flows, can guarantee physically realizable, manifold outputs—but they depend on a high-quality training dataset of already-perfect meshes, which is difficult to scale.
- Real-World Failures: When applied to raw urban data (Nan et al. in PSSNet: Planarity-sensible Semantic Segmentation of Large-scale Urban Meshes), partial scans (Ehm et al. in Geometrically Consistent Partial Shape Matching), or raw sensor point clouds (Wang et al. in Learning local shape descriptors for computing non-rigid dense correspondence), traditional manifold-dependent pipelines fail to ingest the data directly. This limitation even hinders automated testing frameworks, such as discovering physical vulnerabilities in robotic manipulation policies (Goel et al. in Geometric Red-Teaming for Robotic Manipulation).
[Raw / Wild Scan] ---> [Manual / Fragile Repair] ---> [Manifold Mesh] ---> [Downstream Task]
^
THE BOTTLENECK
Strategies for Bypassing the Manifold Constraint
To eliminate this bottleneck, recent research focuses on decoupling the underlying mathematical operators from the explicit connectivity of the input mesh. Four primary strategies have emerged to achieve this.
1. Manifold Proxies and Cages
Instead of repairing a messy mesh directly, why not wrap it in a simplified, mathematically sound container?
This is the intuition behind CageNet: A Meta-Framework for Learning on Wild Meshes. Liu et al. propose enclosing "wild" geometry—complete with disconnected components and non-manifold edges—inside a coarse, watertight manifold "cage." By using differentiable barycentric coordinates, they map data from the messy underlying asset onto the clean cage. Because the cage is a well-behaved manifold, standard geometric neural networks can run on it smoothly, effectively insulating the network from the internal topological chaos of the asset.
A similar proxy strategy is used by Chen et al. in Fast and robust shape diameter function, where thickness is calculated against a simplified "offset surface" rather than the raw, corrupted input mesh, preventing ray-tracing failures.
2. Intrinsic and Unstructured Solvers
Another approach is to redesign numerical solvers so they do not rely on rigid mesh hierarchies. Hierarchical geometric multigrid methods are highly efficient for solving massive linear systems (like those used in physical simulations or shape smoothing), but they traditionally require clean manifold mesh decimation.
To break this dependency, Wiersma et al. in A Fast Geometric Multigrid Method for Curved Surfaces and Liu et al. in Surface Multigrid via Intrinsic Prolongation developed solvers that use intrinsic prolongation operators. By mapping data between resolutions based on local surface geometry rather than rigid 3D distances or explicit face connectivity, these solvers can run directly on unstructured point clouds and non-manifold scans in linear time.
3. Topological Reformulation
Sometimes, the best way to handle a constraint is to change the mathematical language of the problem.
For instance, 3D shape skeletonization algorithms traditionally require a perfectly closed, watertight manifold to define an interior volume. The authors of Coverage Axis: Inner Point Selection for 3D Shape Skeletonization bypass this entirely by reformulating skeletonization as a "Set Cover Problem." By finding the minimum number of inner medial balls required to cover the shape's surface, the algorithm naturally accommodates non-watertight shapes and raw point clouds without needing an explicit definition of an interior volume.
Similarly, for normal orientation, Liu et al. utilize a global winding number property in Fast and Globally Consistent Normal Orientation based on the Winding Number Normal Consistency. Instead of propagating normal directions locally from face to face (which fails at holes or non-manifold junctions), the winding number evaluates a global topological property from a point's perspective, making orientation robust to local defects.
4. Canonical Mappings and Coordinate Spaces
For deep learning architectures that process heterogeneous meshes (objects with varying numbers of vertices and connectivity), researchers use canonical coordinate spaces to abstract away the underlying mesh topology.
In WrappingNet: Mesh Autoencoder via Deep Sphere Deformation, Lei et al. map diverse 3D meshes onto a canonical sphere ("unwrapping") and deform them back into 3D shapes ("wrapping"). This strategy allows the network to learn a shared latent space for cross-category shape interpolation without being constrained by the specific edge-manifold requirements of traditional parameterization techniques (Pentapati et al. in Mesh Compression with Quantized Neural Displacement Fields).
Summary
The shift away from strict manifold dependencies represents a maturation of the geometry processing field. By insulating algorithms from topological defects—whether through proxy cages, intrinsic solvers, or global topological reformulations—modern frameworks can process real-world, "wild" 3D data directly, eliminating the fragile and labor-intensive preprocessing steps of the past.
Key Foundational Work
Reliable implicit surface polygonization using visibility mapping (2006) — contributed 1 concept elements to this direction.
Geodesic Paths on Triangular Meshes (2004) — contributed 1 concept elements to this direction.
Surface reconstruction, one triangle at a time (2004) — contributed 1 concept elements to this direction.