Dynamic SDF (Signed Distance Function)

A Dynamic SDF defines a continuous scalar field over a 4D domain (x, y, z, t). The function f(x, y, z, t) returns the signed distance to the nearest surface at that specific 3D point and moment in time. This continuity is crucial for:

• Temporally coherent rendering without flickering or popping artifacts.
• Enabling smooth interpolation of geometry between discrete timesteps.
• Providing differentiable gradients with respect to both space and time, which is essential for optimization from video data.

Instead of explicitly tracking mesh vertices, motion is encoded within the SDF function itself. This is typically achieved through one of two primary architectures:

• Canonical-to-Observed Deformation: A neural network learns a continuous deformation field T(x_c, t) that maps points from a static canonical 3D space to their observed positions at time t. The SDF is then evaluated in this canonical space: f(T(x, t)).
• Time-Conditioned Networks: A single network directly takes (x, y, z, t) as input and outputs the signed distance for that spatio-temporal coordinate. The network's weights implicitly learn the deformation规律.

The core mathematical property of an SDF—its gradient at any point equals the surface normal at the closest point—extends into the time domain. This allows for the calculation of temporal derivatives that describe how the surface moves.

• The material derivative Dφ/Dt describes the rate of change of the SDF value for a point moving with the surface.
• For a point on the surface (where φ=0), this relates directly to the surface velocity in the normal direction: v_n = - (∂φ/∂t) / ||∇φ||.
• This differentiability enables training via photometric loss from multi-view video, as gradients can flow backwards through time to update the deformation model.

Unlike explicit mesh-based representations, a Dynamic SDF can naturally handle changes in topology without requiring complex re-meshing algorithms. This is critical for modeling dynamic phenomena such as:

• Merging and splitting: Two separate liquid droplets flowing together into one.
• Hole formation: A piece of cloth tearing.
• Phase changes: A solid object melting into a liquid pool. The surface is always defined as the zero-level set { (x,t) | f(x,t)=0 } of the continuous function, which can change its connected components smoothly over time.

A Dynamic SDF collapses the traditional pipeline of 3D reconstruction + scene flow estimation into a single, jointly learned model. The function f(x,t) encodes:

• Instantaneous Geometry: The iso-surface at any fixed t.
• Dense Correspondence: Points with the same canonical coordinates correspond across time.
• Scene Flow: The vector field v(x,t) = ∂T/∂t describing 3D motion. This unified representation ensures consistency between shape and motion, eliminating the need for post-hoc alignment and enabling more accurate temporal interpolation and future frame prediction.

The signed distance value provides direct, queryable information for physical simulation and interaction in dynamic environments.

• Collision Detection: A simple check f(x,t) < 0 determines if a point is inside the object. The magnitude |f(x,t)| gives the penetration depth.
• Contact Normals: The gradient ∇f(x,t) provides the collision normal for response calculations.
• Velocity Coupling: The temporal derivative ∂f/∂t can inform fluid adhesion or friction models. This makes Dynamic SDFs a powerful representation for physics-in-the-loop training and applications in robotics, where understanding contact is essential.

A neural scene representation that extends the static NeRF framework by incorporating time as an input variable. It models a scene's appearance (color) and density as continuous functions of 3D location, viewing direction, and time. Unlike a Dynamic SDF, which explicitly models geometry via distances, Dynamic NeRF uses a volumetric density field, making it highly effective for view synthesis of dynamic scenes but less direct for extracting watertight surfaces.

• Core Mechanism: A multilayer perceptron (MLP) takes (x, y, z, θ, φ, t) and outputs (RGB, σ).
• Primary Use: Generating photorealistic novel views at arbitrary times from monocular or multi-view video.

A continuous, learned vector field that maps points from a observed coordinate at time t back to a shared canonical space. This is the foundational mechanism behind many deformable Dynamic SDF and NeRF models. It separates the learning of a static template shape (in canonical space) from the complex, time-varying motion.

• Function: T(x_observed, t) -> x_canonical.
• Benefit: Enables modeling of non-rigid deformations (like clothing or facial expressions) with a consistent underlying geometry.
• Challenge: Requires regularization to ensure the deformation is bijective (one-to-one and invertible).

The task of estimating a 3D motion vector for every point in a scene between two consecutive times. While a Dynamic SDF provides an implicit surface at each instant, scene flow describes the explicit displacement of those surfaces. The two are deeply connected: the temporal derivative of a Dynamic SDF is related to the scene flow field normal to the surface.

• Output: A dense 3D vector field (u, v, w) for each point (x, y, z).
• Application: Critical for motion prediction, robotic navigation in dynamic environments, and as a supervisory signal for training Dynamic SDFs.

An explicit, point-based representation for dynamic 3D scenes. It models the scene with a set of 3D Gaussians whose attributes—position, rotation, scale, opacity, and spherical harmonic coefficients for color—are defined as continuous functions of time. It contrasts with the implicit representation of a Dynamic SDF, offering extremely fast, real-time rendering suitable for interactive applications.

• Key Innovation: Differentiable splatting rasterization allows for gradient-based optimization from video.
• Advantage: Achieves high-fidelity, real-time performance (> 100 FPS) for dynamic view synthesis, a significant leap over early neural rendering methods.

A modeling strategy where all observations of a deforming object across time are warped into a single, fixed reference frame. A primary SDF or NeRF is then learned in this canonical space. The complexity is offloaded to the deformation field that performs the mapping. This is a common architecture for Dynamic SDFs, as it simplifies learning a coherent shape.

• Process: 1. Estimate deformation from observed space to canonical space. 2. Query the canonical SDF/NeRF. 3. Apply inverse deformation for rendering.
• Benefit: Improves learning efficiency and surface consistency for objects with articulated or periodic motion.

A regularization term used during the training of dynamic scene representations like Dynamic SDFs. It penalizes unrealistic or physically implausible changes in geometry or appearance between consecutive timesteps. This loss is essential for preventing flickering artifacts and ensuring the reconstructed motion is smooth and consistent.

• Common Forms:
  • Penalizing large accelerations in deformation fields.
  • Enforcing similarity between SDF values or colors at a point across adjacent frames after accounting for estimated motion.
• Purpose: Acts as a motion prior, guiding the model when visual evidence is ambiguous or missing.