Particle Bounding Frustum (PBF) for Volumetric Rendering

• Particle Bounding Frustum (PBF) is an analytic per-Gaussian data structure that precisely defines a tight angular frustum for efficient volumetric rendering.
• It computes exact angular and depth bounds from Gaussian parameters and camera extrinsics, significantly reducing unnecessary ray–Gaussian intersection checks.
• Empirical evaluations demonstrate that PBF accelerates rendering (e.g., 327 FPS on MipNeRF360) while enhancing image quality and precision over traditional methods.

A Particle Bounding Frustum (PBF) is an analytic, per-Gaussian data structure introduced to enable exact and efficient volumetric rendering of 3D Gaussians under arbitrary camera models. Each PBF defines a tight viewing frustum in camera space, just enclosing the chosen λ–σ iso-density ellipsoid of a 3D Gaussian, and is fundamental to the 3DGEER method for ray-marching rendering. The PBF enables a one-to-many mapping between camera sub-frustums (CSFs) and particles (Gaussians), dramatically reducing unnecessary ray–Gaussian intersection checks, and is computed purely from the parameters of the Gaussians and camera extrinsics, without projective or sampling approximations (Huang et al., 29 May 2025).

1. Definition and Geometric Interpretation

A 3D Gaussian is parameterized by its mean $\mu \in \mathbb{R}^3$ and covariance $\Sigma \in \mathbb{R}^{3 \times 3}$. Its isodensity surfaces—the locus of points of equal Mahalanobis distance—form ellipsoids,

$(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$

where $\lambda$ specifies the contour (e.g., $\lambda=3$ for a $3\sigma$ ellipsoid). In the angular domain of the camera—parameterizing each ray by its spherical incidence angles $(\theta, \phi)$—this ellipsoid projects to an angular frustum: a prism defined by horizontal $(\theta)$ and vertical $(\phi)$ angular bounds. The PBF is the smallest such frustum tangent to the $\lambda$–ellipsoid, forming a six-plane bounding volume that admits analytic computation.

By associating Gaussians only with those rays whose directions pass through their PBF, the rendering pipeline achieves both exactness and high computational efficiency, markedly improving upon methods using bounding spheres or axis-aligned boxes, which overestimate angular coverage and fail to capture anisotropy.

2. Mathematical Derivation and Construction

The construction of the PBF for a Gaussian involves several analytic steps:

1. Transformation to Camera Frame: The Gaussian's mean and covariance are transformed to the camera coordinate system using the camera extrinsics $\Sigma \in \mathbb{R}^{3 \times 3}$0, with

$\Sigma \in \mathbb{R}^{3 \times 3}$1

where $\Sigma \in \mathbb{R}^{3 \times 3}$2 encodes ellipsoid scaling.

2. Euler Angle Parameterization: Rays are parameterized by their incidence angles:

$\Sigma \in \mathbb{R}^{3 \times 3}$3

where $\Sigma \in \mathbb{R}^{3 \times 3}$4 is the 3-vector direction.

3. Frustum Planes: The angular bounds are given by four planes in homogeneous coordinates:

$\Sigma \in \mathbb{R}^{3 \times 3}$5

$\Sigma \in \mathbb{R}^{3 \times 3}$6

4. Tangency Condition: Each frustum plane is required to be tangent to the $\Sigma \in \mathbb{R}^{3 \times 3}$7–ellipsoid. The tangent planes satisfy:

$\Sigma \in \mathbb{R}^{3 \times 3}$8

under the appropriate transformation, with $\Sigma \in \mathbb{R}^{3 \times 3}$9 for $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$0. The angular bounds are then given by the real roots $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$1, $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$2 of the quadratics:

$(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$3

with $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$4.

5. Depth Planes: Near and far planes along the z-axis in camera frame are given by

$(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$5

enclosing the ellipsoid's extremal z-values.

If the angular quadratic's discriminant is negative (camera center inside the ellipsoid), angular bounds default to camera's field-of-view.

3. Algorithmic Workflow and Complexity

The computation of PBFs is performed independently for each Gaussian, with algorithmic steps:

1. Transform $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$6 and $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$7 to camera coordinates.
2. Form the $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$8 matrix $(x-\mu)^\top \Sigma^{-1}(x-\mu) = \lambda^2,$9 via $\lambda$0.
3. Solve quadratics for angular bounds $\lambda$1, $\lambda$2.
4. Compute $\lambda$3.
5. Package bounds as six-plane frustum.

This algorithm requires only $\lambda$4 operations per frame for $\lambda$5 Gaussians, with a per-Gaussian computation cost that is approximately an order of magnitude less than building bounding volume hierarchies or deploying sampling-based approximations.

Pseudocode (Gaussian-wise):

$\lambda=3$1

4. Ray–Gaussian Association and Rendering

The PBF facilitates fast and exact ray–Gaussian association. The image is partitioned into camera sub-frustums (CSFs), each defined by angular and depth intervals. An association is made by checking for overlap between the PBF intervals $\lambda$6 and the CSF bounds.

Table: Comparison of Ray–Particle Culling Methods | Method | Angular Fit | Algorithmic Overhead | |----------------|----------------------|-----------------------| | PBF | Tight, anisotropic | O(N), analytic | | Spherical Bound| Overestimated | O(N), simple | | AABB in angle | Overdrawn | O(N), suboptimal |

During rendering, each CSF only processes its associated Gaussians, sorting them by depth and marching rays precisely through their contributions. This approach yields zero false negatives and minimal false positives, outperforming prior methods that rely on bounding spheres (ignoring anisotropy) or axis-aligned bounding boxes (causing angular overdraw). Early ray terminates are triggered when accumulated opacity reaches unity.

5. Empirical Performance and Qualitative Gains

Empirical evaluations on datasets such as ScanNet++ and MipNeRF360 demonstrate the superiority of the PBF in terms of rendering quality and efficiency. On ScanNet++, PBF–CSF association achieves the lowest LPIPS (0.1245) and smallest cross-method LPIPS gap ($\lambda$7) compared to EWA (0.1250/$\lambda$8) and UT (0.1249/$\lambda$9). In run-time profiling, PBF association requires 3.98 ms/frame, compared to 8.68 ms for EWA+SnugBox and 4.38 ms for UT+SnugBox, corresponding to a 2.5–5× speedup.

On MipNeRF360, the PBF-enabled 3DGEER renders at 327 FPS, outperforming other exact ray-marchers such as EVER (36 FPS) and 3DGRT (52–68 FPS), while matching the speed of approximate Gaussian splatting (3DGS at 343 FPS). PBF incurs no loss in reconstruction quality, instead improving PSNR by 0.2–0.5 dB and LPIPS by 0.004–0.01 relative to projective approximations (Huang et al., 29 May 2025).

6. Advantages and Distinctive Features

The PBF provides several distinct advantages:

• Analytic, per-particle representation, with exact angular and depth bounds.
• Captures particle anisotropy, unlike isotropic bounding spheres.
• Minimal per-frame computational overhead ($\lambda=3$0, small constant), obviating the need for complex data structures or sampling.
• Enables real-time, exact, and state-of-the-art volumetric rendering with empirical benefits in both speed and image quality.

The use of PBF eliminates the assumptions of Taylor-series or unscented transform approximations, providing exactness with no quality degradation on a broad range of camera models and field-of-view scenarios.

7. Context within Differentiable Rendering and Future Implications

The introduction of the PBF within the 3DGEER pipeline addresses the primary limitations of 3D Gaussian Splatting (3DGS). While 3DGS achieves real-time speeds by projecting 3D Gaussians as 2D image-space Gaussians, this approximation degrades at large field-of-view or with strong scene anisotropies. PBF-based approaches not only eliminate these projective limitations but also deliver computational speed comparable to splatting and higher than other exact volumetric rendering frameworks.

This suggests that analytic, per-particle culling structures such as PBF could become the standard for exact real-time volumetric rendering in neural graphics pipelines, particularly as 3D Gaussians are increasingly adopted for scene representation in vision and graphics (Huang et al., 29 May 2025).