Volume Rendering Integral

Transmittance represents the fraction of light that survives (is not absorbed or scattered) along a segment of the ray from the camera to a given sample point t. It is an exponential decay function of the volume density σ integrated along the path.

• Mathematical Role: Acts as an attenuation weight for radiance contributions from deeper points.
• Key Property: Monotonically decreases from 1 (at the camera) towards 0 as the ray penetrates dense matter.
• In NeRF: The neural network predicts density σ, which is used to compute transmittance via numerical quadrature (e.g., the cumprod operation).

Volume density (σ, sigma) is a scalar field defining the differential probability of a ray interacting with the medium (via absorption or out-scattering) per unit length at a point in space. It is the core geometric signal in a radiance field.

• Units: Inverse length (e.g., m⁻¹).
• Defines Opacity: High density values lead to rapid light attenuation (high opacity), while low values create transparent regions.
• In NeRF: The MLP outputs σ as a function of 3D location (x, y, z), implicitly defining the scene's surfaces where density is locally high.

Emitted radiance (c) is the spectral color (typically RGB) emitted or reflected from an infinitesimal volume element at a point along the ray, before being attenuated by the medium in front of it.

• View-Dependence: In models like NeRF, this color is a function of both 3D location (x, y, z) and viewing direction (θ, φ), allowing the representation of non-Lambertian effects like specular highlights.
• Physical Interpretation: Represents the combined effect of intrinsic albedo, local lighting, and material BRDF.
• Rendering Role: The raw color contribution that gets weighted by transmittance and density.

The continuous integral is approximated using numerical quadrature, typically the composite quadrature rule. The ray is partitioned into N segments, and the integral is evaluated as a weighted sum of samples.

• Core Procedure: C = Σ (T_i * (1 - exp(-σ_i δ_i)) * c_i)
• Alpha Compositing: The term (1 - exp(-σ_i δ_i)) acts as the alpha value in classic alpha blending, representing the opacity of the i-th segment.
• Implementation: This discretization is fully differentiable, enabling gradient-based optimization of the density σ and color c fields from 2D images.

The ray is defined by an origin o and direction d, parameterized by a scalar t: r(t) = o + t*d. The integral is evaluated between the near bound t_n and far bound t_f.

• Bounds Selection: Defines the volume of interest. In unbounded scenes, techniques like inverted sphere parameterization or contracting the space are used.
• Hierarchical Sampling: Efficient rendering (as in NeRF) uses a two-stage process: a coarse network to estimate density distribution, guiding importance sampling for a fine network.
• Physical Basis: The parameter t has units of distance, tying the abstract integral to the physical scale of the scene.

The Volume Rendering Integral generalizes several classic computer graphics equations:

• Surface Rendering: In the limit of an infinitely thin surface (a Dirac delta in density), the integral reduces to the standard rendering equation for surfaces.
• Alpha Blending: The discretized form is mathematically identical to back-to-front alpha compositing, providing a clear link to rasterization pipelines.
• Participating Media: It directly models effects like fog, smoke, and translucent materials where light undergoes multiple scattering events. Advanced extensions (e.g., Neural Reflectance Fields) add terms for in-scattered light from other directions.

Ray marching is the iterative numerical algorithm used to approximate the volume rendering integral. It works by:

• Casting a ray from the camera through each pixel.
• Taking discrete, adaptive steps along the ray.
• At each sample point, querying the volumetric properties (density, color).
• Accumulating these samples using the alpha compositing equation. This process is the computational engine for rendering Neural Radiance Fields (NeRF) and Signed Distance Functions (SDFs). Advanced variants like sphere tracing are used for efficient SDF rendering.

Alpha compositing is the discrete, practical implementation of the volume rendering integral used in computer graphics. For a ray with sampled points, it calculates the final pixel color C as: C = Σ (T_i * α_i * c_i) where:

• α_i is the opacity (from density) at sample i.
• c_i is the emitted color at sample i.
• T_i is the transmittance, the fraction of light that reaches sample i, calculated as Π (1 - α_j) for all previous samples j. This over operator provides the foundational math for blending semi-transparent layers in volume rendering and neural graphics.

A participating medium is any material that affects light via absorption, scattering, or emission as it travels through space. The volume rendering integral models light transport through such media. Key phenomena include:

• Absorption: Light is converted to other energy (e.g., heat).
• Out-scattering: Light is deflected away from the ray path.
• In-scattering: Light from other directions is deflected into the ray path.
• Emission: The medium itself emits light (e.g., a flame). In neural graphics, a NeRF treats empty space and objects as a continuous participating medium, assigning each 3D point a volume density (for absorption/out-scattering) and an emitted radiance (color).

The Radiative Transfer Equation (RTE) is the fundamental differential equation describing the change in radiance along a path through a participating medium. It is the physical law from which the volume rendering integral is derived. The RTE states that the change in radiance L at a point is due to:

• Losses: Extinction (absorption + out-scattering).
• Gains: Emission + in-scattering from all directions. The volume rendering integral is the formal solution to the RTE along a ray, assuming no in-scattering (the null-scattering or emission-absorption model). Full RTE solvers are used in high-fidelity physics simulations like climate modeling and astrophysics.

Transmittance T(t) is a key term within the volume rendering integral, representing the fraction of light that travels from the ray origin to a distance t without being absorbed or scattered. It is defined as: T(t) = exp( -∫_0^t σ(s) ds ) where σ(s) is the extinction coefficient (density) at point s.

• In alpha compositing, transmittance is calculated cumulatively from sampled opacities.
• It is the mathematical expression of visibility or attenuation through fog, smoke, or a neural density field. High transmittance means most light passes through; low transmittance means the medium is opaque.

The emission-absorption model is the specific light transport model underpinning the standard volume rendering integral and neural radiance fields. It simplifies the full Radiative Transfer Equation (RTE) by considering only:

1. Emission: The medium emits light of a certain color at each point.
2. Absorption/Extinction: The medium attenuates light traveling through it. It explicitly neglects in-scattering (light bouncing into the ray from other directions). This simplification is computationally tractable and sufficient for modeling many phenomena, including the appearance of clouds, smoke, and the opaque surfaces represented in a NeRF, where surface-like behavior emerges from regions of very high density.