Imagine trying to draw a crisp, 90-degree corner using only a thick, blunt marker. No matter how carefully you position the tip, the physical width of the marker will naturally round out the sharp point.
In 3D geometry processing, standard surface reconstruction pipelines suffer from a digital equivalent of this phenomenon, known as geometric over-smoothing. When algorithms process raw 3D data—such as point clouds from LiDAR or depth maps from consumer sensors—they face a fundamental challenge: distinguishing between random sensor noise and high-frequency geometric details like edges, creases, and corners.
As noted by HyeokHyen Kwon et al. in Data-driven depth map refinement via multi-scale sparse representation, conventional noise reduction filters often wash out critical structural information because they lack the capacity to differentiate a sharp physical boundary from an erroneous data spike. The result is a persistent bias toward rounded, "bloaty" reconstructions that fail to accurately reflect the true geometry of man-made or mechanical objects.
The Root Cause: The Low-Pass Bias of Smoothness Priors
To understand why sharp features degrade, we must look at the mathematical objectives driving traditional reconstruction. Most pipelines rely on global smoothness priors or continuous implicit functions to resolve ambiguities in noisy data.
For instance, in multi-view stereo or global surface fitting, algorithms frequently employ optimization objectives that penalize rapid changes in surface orientation. As Michael Goesele observed in Multi-View Stereo Revisited, while global smoothness priors are highly effective at filling holes where data is missing, they inherently act as low-pass filters. They treat sharp geometric discontinuities as anomalies to be smoothed away.
This issue is compounded by standard point-based loss functions used in modern deep learning models. As explained by Weijia Wang et al. in Noise4Denoise: Leveraging noise for unsupervised point cloud denoising, unsupervised denoising models often prioritize global surface continuity. When a model optimizes for a smooth manifold, the mathematical step-function of a sharp, acute corner incurs a massive penalty. This forces the network to produce rounded approximations where precise geometric intersections are required.
Furthermore, even if a neural network perfectly learns a sharp geometry implicitly, the post-processing step used to extract an explicit mesh often ruins it. Grid-based extraction methods like Marching Cubes sample continuous fields at fixed, regular intervals. As Nissim Maruani et al. point out in VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams, these grid-based approaches fail to capture high-frequency information, effectively blurring discrete feature intersections into continuous, rounded approximations during the final meshing phase.
Breaking the Smoothness Bias with Dual and Voronoi Representations
To preserve sharp features, researchers have shifted away from uniform grid sampling toward representations that can adaptively snap to geometric discontinuities.
One prominent approach is leveraging Dual Contouring rather than Marching Cubes. While Marching Cubes places vertices strictly along the edges of grid cells (cutting across corners and truncating them), Dual Contouring generates vertices inside the grid cells. This allows a single vertex to be positioned exactly at the intersection point of multiple surface normals.
In Neural Dual Contouring (NDC), Zhiqin Chen et al. train a data-driven framework to predict these internal vertex positions and edge crossings explicitly. By learning from data rather than relying on fixed grid interpolations, the network learns to snap vertices directly onto sharp mechanical edges and corners, producing high-fidelity meshes with significantly fewer triangles.
Marching Cubes (Isosurface cuts across the corner)
| /
----+-----*-----
| / | * = Generated Vertex
| / | (The corner is truncated and rounded)
----*-------+-----
/ | |
Dual Contouring (Vertex placed at the feature intersection)
| |
----+-------+-----
| * | * = Generated Vertex
| / \ | (The vertex snaps to the true corner
----+-/---v-+----- inside the cell)
|/ |
Alternatively, instead of restricting the geometry to a cubical grid, the surface can be defined by cellular structures that naturally adapt to boundaries. The authors of VoroMesh demonstrate this by optimizing the positions of Voronoi diagram generators. Because the boundaries between Voronoi cells are mathematically defined by sharp, intersecting planes, optimizing these generators allows the resulting mesh to form perfectly sharp, watertight intersections without the aliasing or over-smoothing typical of grid methods.
Structural Guidance and Primitive Assembly
Another powerful strategy involves guiding the reconstruction process using explicit geometric fields or structural building blocks rather than unconstrained optimization.
Instead of letting points freely drift during upsampling or denoising, algorithms can use directional guidance fields to align points along structural lines. For example, in iPUNet: Iterative Cross Field Guided Point Cloud Upsampling, Hao Pan et al. introduce self-supervised cross fields. These fields act as a geometric compass, guiding the densification of the point cloud so that new points are explicitly pulled toward crisp edges and corners rather than scattering randomly and blurring the geometry.
Taking this structural logic to its conclusion, some methods abandon dense mesh deformation entirely when dealing with man-made objects. In Structure-Aware Surface Reconstruction via Primitive Assembly, Jingen Jiang et al. propose reconstructing surfaces by assembling geometric primitives like cylinders, spheres, and planes. Because these primitives are mathematically exact, their intersections naturally form perfectly sharp, clean edges. By framing reconstruction as a primitive assembly problem, the authors bypass the over-smoothing traps of dense implicit functions entirely, ensuring that the structural logic of the original object is preserved.