Surface Reconstruction

Moving Least Squares is a projection-based method that defines a smooth, continuous surface directly from the raw point cloud without explicit meshing. Each point is projected onto a local polynomial approximation (often a plane or quadric) fitted to its neighbors.

• Core Function: Provides a continuous surface definition and resampling for point cloud smoothing and up-sampling.
• Key Feature: Effectively reduces noise and fills small holes while preserving the underlying surface detail.
• Two Types: MLS projection (for surface definition) and MLS up-sampling (for point cloud refinement).
• Application: Preprocessing step for other reconstruction methods, or for direct rendering of point clouds as surfaces.

Volumetric methods discretize space into a voxel grid (or octree) and classify each voxel as inside, outside, or on the surface. The Marching Cubes algorithm then extracts a triangle mesh by constructing polygons within voxels that intersect the estimated surface.

• Pipeline: Point cloud → convert to volumetric representation (e.g., Signed Distance Field) → apply Marching Cubes.
• Flexibility: Can integrate data from multiple scans and handle complex topology naturally.
• Trade-off: Resolution is limited by voxel size; high detail requires large memory (mitigated by adaptive octrees).
• Ubiquity: The de facto standard for extracting meshes from medical CT/MRI scans and implicit functions.

A point cloud is the primary raw input for surface reconstruction. It is a discrete set of data points in a 3D coordinate system, representing the external surfaces of objects as captured by sensors like LiDAR, structured light, or depth cameras. Each point contains XYZ coordinates and may include color or intensity data. Surface reconstruction algorithms 'skin' this sparse, noisy data to infer the continuous manifold that connects the points.

A Signed Distance Field (SDF) is a powerful implicit surface representation used in modern reconstruction. It defines a continuous volumetric function where the value at any 3D point is its shortest signed distance to the inferred surface (negative inside, positive outside). The surface is defined at the zero-level set of this function. Neural networks, like those in DeepSDF, can learn this function, enabling high-fidelity reconstruction from incomplete scans.

A voxel grid is a discrete, volumetric representation of 3D space, analogous to a 3D pixel grid. Each voxel (volume element) can store occupancy probability, SDF value, or color. Surface reconstruction often involves converting a point cloud into a voxel grid, then applying algorithms like Marching Cubes to extract a polygonal mesh from the voxel data. This representation is computationally intensive but compatible with 3D convolutional neural networks.

A polygon mesh is the most common output of surface reconstruction. It represents a 3D surface as a collection of vertices, edges, and faces (typically triangles or quads). This explicit representation is lightweight and directly usable by graphics engines, CAD software, and 3D printers. Reconstruction algorithms like Poisson Surface Reconstruction and Ball-Pivoting generate watertight meshes by connecting neighboring sample points into a coherent surface topology.

Poisson Surface Reconstruction is a seminal algorithm for creating smooth, watertight surfaces from oriented point clouds. It formulates reconstruction as solving a Poisson equation, treating the input points as samples of an indicator function whose gradient is a vector field defined by the point normals. The resulting surface is the iso-surface of the solved function, producing robust meshes even with noisy data and missing regions.

Marching Cubes is a fundamental computer graphics algorithm for extracting a polygonal mesh of an iso-surface from a 3D scalar field (like an SDF or voxel occupancy grid). It 'marches' through the volume cube-by-cube, determining the mesh topology within each cube based on the field values at its eight corners. It is the standard method for converting implicit surface representations (SDFs) into explicit triangle meshes for rendering and simulation.