Signed Distance Function (SDF)

The core value of an SDF, f(p), at a 3D point p is its Euclidean distance to the nearest surface. The sign is the critical differentiator:

• Positive (> 0): Point is outside the object.
• Zero (= 0): Point lies exactly on the surface.
• Negative (< 0): Point is inside the object. This sign provides an implicit, unambiguous definition of interior vs. exterior, which is essential for tasks like collision detection and mesh Boolean operations.

A valid SDF satisfies the Eikonal equation: ||∇f(p)|| = 1 almost everywhere. This means the magnitude of its gradient is 1, indicating the function's value changes at a rate of 1 unit per unit of movement toward the surface.

• Implication: The gradient ∇f(p) at any point is a unit vector pointing in the direction of the nearest surface point. This property enables efficient algorithms like ray marching and provides surface normals for rendering and physics (normal = ∇f(p)).

An SDF is a 1-Lipschitz continuous function. This means the absolute difference in its value between any two points is at most the Euclidean distance between them: |f(p) - f(q)| ≤ ||p - q||.

• Consequence: The function is well-behaved and does not change arbitrarily quickly. This mathematical stability is crucial for numerical optimization and learning-based reconstruction, as it provides a smooth, predictable loss landscape for gradient-based methods.

Complex geometry can be constructed through CSG (Constructive Solid Geometry) operations by combining simple SDFs using min/max functions.

• Union: union_sdf(p) = min( sdf_A(p), sdf_B(p) )
• Intersection: intersection_sdf(p) = max( sdf_A(p), sdf_B(p) )
• Difference (A - B): difference_sdf(p) = max( sdf_A(p), -sdf_B(p) ) These operations are exact and differentiable, making SDFs ideal for procedural modeling and differentiable rendering pipelines.

An SDF is continuously differentiable (C1) everywhere except at the medial axis (the set of points equidistant to two or more surface points). This smooth gradient field is a key advantage over discrete representations (voxels, meshes).

• Applications:
  • Physics & Simulation: Enables efficient collision response and fluid interaction.
  • Inverse Problems: Allows gradient-based optimization of shape from images (e.g., in differentiable rendering).
  • Neural Networks: Can be learned directly as a neural implicit representation (e.g., DeepSDF, NeRF with SDF).

SDFs offer distinct trade-offs compared to common 3D data structures.

• vs. Mesh/Polygon Soup: SDFs are watertight by definition, avoiding non-manifold edges. They are infinite-resolution but require evaluation, not direct storage of vertices.
• vs. Voxel Grid: SDFs are continuous and memory-efficient, avoiding the cubic memory growth (O(n³)) of dense voxel grids. They suffer from no discretization aliasing.
• vs. Point Cloud: SDFs define a complete surface with interior/exterior, whereas point clouds are an unstructured set of surface samples with no inherent connectivity or occupancy information.

SDFs are the core data structure for ray marching, a rendering technique used to visualize complex implicit surfaces in real time. The algorithm efficiently finds surface intersections by stepping along a ray according to the SDF's distance value, guaranteeing a hit. This is fundamental to ShaderToy-style graphics, volumetric effects, and real-time global illumination for procedurally defined scenes.

• Key Advantage: Provides an analytic surface test, enabling effects like soft shadows, ambient occlusion, and accurate reflections directly from the distance field.
• Example: The demoscene and tools like ShaderToy heavily rely on SDF primitives and operations to create intricate 3D visuals entirely from code.

In physics engines and robotic simulation, SDFs provide a highly efficient method for proximity queries and collision detection. The value at any point instantly indicates penetration depth, which is directly usable for calculating collision responses and contact forces.

• Key Advantage: The gradient of the SDF is the surface normal, providing collision resolution data (penetration depth and direction) in a single query.
• Use Case: NVIDIA PhysX and other simulators use SDFs (or distance fields) for fast, robust collision handling between complex, deforming shapes. In robotics, SDFs of the environment enable efficient motion planning by quantifying obstacle proximity.

SDFs are the canonical output representation for many modern neural implicit reconstruction techniques. Instead of storing a mesh, a neural network (e.g., an MLP) is trained to map 3D coordinates to signed distance values, creating a continuous, memory-efficient model of an object or scene.

• Key Advantage: Provides a continuous, differentiable surface representation ideal for optimization from 2D images or 3D point clouds.
• Example: Architectures like DeepSDF and Occupancy Networks learn to encode a shape's SDF, enabling high-fidelity shape completion, interpolation, and generation from partial data.

In robotics, an SDF of the environment is a critical map representation for motion planning. It allows a planner to efficiently compute the distance and gradient to the nearest obstacle for any proposed robot configuration, enabling gradient-based optimization for safe, smooth paths.

• Key Advantage: Enables gradient-based optimization for trajectory planning. Algorithms can follow the SDF's gradient away from obstacles to minimize collision cost.
• Use Case: Frameworks like CHOMP (Covariant Hamiltonian Optimization for Motion Planning) use pre-computed world SDFs to optimize trajectories that maximize clearance from obstacles while satisfying dynamic constraints.

SDFs enable elegant and robust Constructive Solid Geometry (CSG). Complex shapes can be built by combining primitive SDFs (spheres, boxes) using Boolean operations (union, intersection, difference) defined with simple min/max functions on their distance values.

• Key Advantage: CSG operations on SDFs are artifacts-free and mathematically clean, unlike polygonal mesh Booleans which often suffer from precision issues.
• Example: The Inigo Quilez articles on SDFs demonstrate how to model intricate shapes for rendering using these compositional principles. This is foundational for procedural modeling and industrial CAD in research contexts.

The zero-level set of an SDF defines a precise, watertight surface. Algorithms like Marching Cubes or Dual Contouring can be run on a discretized SDF (a voxel grid where each voxel stores a distance value) to extract a high-quality polygonal mesh.

• Key Advantage: Guarantees watertight, manifold meshes from any implicit representation, which is crucial for 3D printing and simulation.
• Use Case: This is the final step in many 3D reconstruction pipelines. After a neural network learns an implicit SDF, Marching Cubes is applied to extract a usable mesh for downstream applications in graphics and engineering.

A neural implicit representation encodes a 3D shape or scene using a neural network (typically a Multi-Layer Perceptron) that maps 3D spatial coordinates to a continuous field value, such as occupancy probability or signed distance. Unlike discrete voxel grids, this provides a memory-efficient, infinitely detailed model. It is the underlying architecture for modern techniques like Neural Radiance Fields (NeRF) and many learned SDF models, enabling high-fidelity reconstruction and novel view synthesis from sparse inputs.

A voxel grid is a discrete, volumetric representation where 3D space is partitioned into a regular lattice of cubic volume elements called voxels. Each voxel stores attributes like binary occupancy, density, or color.

• Key Contrast with SDFs: Voxel grids are discrete and memory-intensive at high resolutions, whereas SDFs are continuous functions.
• Common Use: Foundational for 3D convolutional neural networks (3D CNNs) in early deep learning for 3D data. Modern sparse voxel and hierarchical representations help mitigate the cubic memory growth.

An occupancy network is a specific type of neural implicit representation where the network learns to predict a probabilistic occupancy field. For any 3D query point, the network outputs a value between 0 and 1, representing the likelihood that the point is inside an object.

• Difference from SDF: It models a classification (inside/outside) rather than a regression to a precise distance metric.
• Application: Widely used for 3D reconstruction from point clouds or single images, and is a core component in occupancy prediction for autonomous driving to reason about unknown or dynamic space.

A point cloud is a set of discrete data points in a 3D coordinate system, representing the external surface of an object or scene. It is the raw output of sensors like LiDAR and RGB-D cameras.

• Relation to SDF: Point clouds are sparse and surface-only. SDFs can be generated from point clouds via reconstruction algorithms (e.g., Poisson reconstruction) or learned directly from them using neural networks.
• Characteristics: Unstructured, lacks connectivity, and requires processing (e.g., normal estimation, meshing) for many downstream robotics tasks like collision checking.

Surface reconstruction is the process of creating a continuous 2D manifold surface (typically a triangle mesh) from discrete 3D data, such as a point cloud or an SDF.

• The Marching Cubes Algorithm: The canonical algorithm for extracting a mesh iso-surface from a volumetric SDF or occupancy grid. It 'marches' through voxels, generating triangles based on the field values at the cube's vertices.
• Pipeline: SDFs are often the intermediate, continuous representation that is then passed to a surface reconstruction algorithm to produce a usable mesh for simulation, rendering, or manufacturing.

Differentiable rendering is a framework that allows gradients to be computed from a rendered 2D image back to the underlying 3D scene parameters (geometry, appearance, lighting). This enables the optimization of 3D representations using only 2D image supervision.

• Critical for Learned SDFs: Modern methods for learning SDFs from images (e.g., NeuS) rely on differentiable renderers. The SDF defines the geometry, and a rendering equation is applied to produce a 2D image; the difference from a real image generates a loss that propagates back to refine the SDF.
• Bridges Graphics and Vision: It is the key technology enabling optimization-based 3D reconstruction without explicit 3D supervision.