TetraSDF: Analytic Isosurface Extraction with Multi-resolution Tetrahedral Grid¶

Conference: ECCV 2026
Paper: CVF Open Access
Code: https://seonghunn.github.io/tetrasdf/
Area: 3D Vision
Keywords: neural implicit surface, signed distance function, analytic isosurface extraction, continuous piecewise affine (CPWA), tetrahedral grid

TL;DR¶

TetraSDF introduces a continuous piecewise affine (CPWA) neural SDF framework combining a multi-resolution tetrahedral grid with a shallow ReLU MLP, leveraging the inherently affine nature of barycentric interpolation to achieve exact analytic zero-level set isosurface extraction while preserving high-frequency shape fitting capability.

Background & Motivation¶

In neural implicit surface modeling and geometry processing, extracting explicit triangle meshes from continuous scalar fields, such as neural Signed Distance Functions (SDFs), serves as a fundamental interface for downstream rendering, physics simulations, and CAD pipelines. For decades, classical isosurfacing techniques, including Marching Cubes, Marching Tetrahedra, and Dual Contouring, have converted implicit fields into meshes via discrete grid sampling. However, these methods inevitably suffer from sampling-dependent discretization errors and staircase artifacts; significantly suppressing the discrepancy between the extracted mesh and the ground-truth zero-level set requires inflating the grid resolution, which leads to massive triangle counts and severe over-fragmentation.

To overcome the inherent limitations of sampling approximations, analytic meshing paradigms such as Analytic Marching interpret ReLU MLPs as continuous piecewise affine (CPWA) functions. By computing the intersections between the folded decision hyperplanes of ReLU neurons and spatial polyhedral cells, they analytically extract the exact zero-level set of the network. Nonetheless, relying solely on plain ReLU MLPs severely restricts geometric expressiveness due to spectral bias, making it extremely difficult to capture intricate high-frequency details. Subsequent grid-based positional encoders, such as Instant-NGP (HashGrid), excel at high-frequency learning but fundamentally depend on trilinear interpolation. Because trilinear interpolation introduces higher-order polynomial terms that break the CPWA property, recent attempts like TropicalNeRF must rely on localized eikonal assumptions and heuristic diagonal-plane slicing, failing to provide strict mathematical guarantees that the extracted mesh matches the network's zero-level set.

Bridging the gap between high-frequency grid representations and exact analytic isosurface extraction requires a fundamental rethinking of spatial interpolation primitives. The core idea is to build a multi-resolution spatial encoder utilizing the strictly affine nature of barycentric interpolation over tetrahedral grids, ensuring that the combined encoder-ReLU-MLP architecture remains globally continuous piecewise affine (CPWA) and deriving an analytic input preconditioner to eliminate directional metric bias for exact, high-fidelity zero-level set mesh extraction.

Method¶

Overall Architecture¶

The pipeline of TetraSDF operates in three cohesive stages: First, an analytic linear preconditioner transforms input coordinates to neutralize the intrinsic metric anisotropy of the tetrahedral grid decomposition. Second, coordinates are processed by a multi-resolution tetrahedral positional encoder that performs barycentric feature interpolation across resolution levels, feeding concatenated features into a compact ReLU MLP to predict the continuous signed distance scalar. Third, an analytic extraction stage constructs an initial polyhedral complex skeleton and performs grid-aware edge subdivision by jointly tracking tetrahedral boundaries and ReLU hyperplanes via barycentric masks and region indicators, ultimately recovering an exact zero-level set triangle mesh.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input 3D Coordinates x"] --> B["Analytic Input Preconditioner<br/>metric whitening to eliminate anisotropy"]
    B --> C["Multi-resolution Tetrahedral Grid & Barycentric Interpolation<br/>preserves global CPWA feature mapping"]
    C --> D["Shallow ReLU MLP<br/>predicts continuous scalar SDF"]
    D --> E["Tensorized Initial Skeleton & Grid-aware Edge Subdivision<br/>jointly tracks polyhedral cells & neuron hyperplanes"]
    E --> F["Analytic Zero-level Set Mesh Extraction<br/>discretization-error-free triangle mesh"]

Key Designs¶

1. Multi-resolution Tetrahedral Grid & Barycentric Interpolation: Preserving the Global CPWA Property Standard grid-based positional encoders (such as HashGrid) utilize trilinear interpolation over cubic voxels, which yields cubic polynomials in coordinates and fundamentally disrupts spatial piecewise affinity. To prevent spectral bias while preserving global CPWA structure, TetraSDF introduces an $L$-level multi-resolution tetrahedral grid. At each level $\ell \in \{0, \dots, L-1\}$, the grid resolution scales geometrically as $N_\ell = \lfloor N_{\min} \gamma^\ell \rfloor$. The unit cube is uniformly subdivided into $N_\ell^3$ cubic cells, each decomposed into six congruent tetrahedra according to a consistent topological scheme. For any query point $x$, it is contained within a specific tetrahedron $T^{(\ell)}(x)$ whose four vertices map to learnable feature vectors via a spatial hash function. Because barycentric coordinates $w^{(\ell)}_v(x)$ inside a tetrahedron are strictly affine in $x$: $$w^{(\ell)}_v(x) = J_T x + b_T, \quad \sum_{v \in V(T^{(\ell)}(x))} w^{(\ell)}_v(x) = 1, \quad w^{(\ell)}_v(x) \ge 0$$ the interpolated feature vector within each tetrahedron is affine in $x$. Since concatenations of affine functions remain affine, the concatenated features across all resolution levels are strictly affine over each polyhedral cell $\mathcal{C}_x = \bigcap_{\ell=0}^{L-1} T^{(\ell)}(x)$. When composed with a subsequent ReLU MLP—which is itself piecewise affine—the entire tetrahedral network $f(x) = \nu^{(M)}(\tau(x))$ defines a globally continuous piecewise affine (CPWA) field.

2. Tensorized Initial Skeleton & Grid-aware Edge Subdivision: Joint Topological Cell Tracking Directly enumerating all intersecting multi-resolution polyhedral cells $\mathcal{C}_x$ is computationally intractable, and conventional edge subdivision algorithms lack spatial grid awareness. TetraSDF resolves this through a GPU-parallel tensorized skeleton construction coupled with joint cell-neuron topology tracking. First, observing that the six-tetrahedra decomposition across all resolution levels shares exactly six fixed plane normals with varying level-dependent offsets, the algorithm extracts polyhedral vertices (intersections of $\ge 3$ independent planes) and edges (intersections of $\ge 2$ non-parallel planes) via parallel tensor operations, establishing the initial skeleton $(V, E)$. Next, to jointly determine spatial boundary locations, the method introduces levelwise barycentric masks $m^{(\ell)}(x) \in \{0, 1\}^4$ (where $w_i^{(\ell)}(x) \le \epsilon_b$ sets the entry to 0, directly indicating whether $x$ lies on a face, edge, or vertex) and a global region indicator $r(x) = \bigoplus_{\ell=0}^{L-1} [a^{(\ell)}(x), t^{(\ell)}(x)] \in \mathbb{Z}^{4L}$ (uniquely indexing the anchor voxel corner and tetrahedron offset). When traversing the folded decision hyperplanes $H_k^{(m)}$ of ReLU neurons, sign vectors $s_k^{(m)}(x)$ are paired with grid region indicators, and local perturbations over zero entries identify adjacent polyhedral cells and linear regions. For any edge $(x_0, x_1)$ crossing a decision boundary, its exact zero-crossing is calculated via linear interpolation: $$\hat{x}_{0,1} = (1-w)x_0 + w x_1, \quad w = \frac{|\nu_k^{(m)}(x_0)|}{|\nu_k^{(m)}(x_0)| + |\nu_k^{(m)}(x_1)|}$$ producing an exact candidate edge and vertex set in parallel on the GPU.

3. Analytic Input Preconditioner: Mitigating Geometric Anisotropy and Directional Bias Decomposing a cube into six congruent tetrahedra causes the local Jacobian $J_T = \frac{\partial w}{\partial x}$ of barycentric weights to exhibit strong spatial anisotropy. In backward feature updates, this leads to non-uniform effective learning rates across different spatial orientations. Specifically, the average local Riemannian metric $M = \frac{1}{6} \sum_{T} J_T^\top J_T$ disproportionately amplifies directions orthogonal to the cube's main body diagonal, inflating the spectral condition number to 16.39 and introducing noticeable directional bias when fitting high-curvature geometries. TetraSDF derives a closed-form analytic linear preconditioner $A^*$ applied directly to the inputs, $x' = A^* x$. By whitening the encoder-induced metric tensor, $A^*$ reduces the spectral condition number from 16.39 to 5.05. This stabilizes gradient descent, enforces isotropic learning behavior, and markedly improves geometric accuracy along sharp edges and curved features.

Loss & Training¶

The network is optimized end-to-end using a weighted combination of a narrow-band $\ell_1$ distance loss and an eikonal regularizer: $$\mathcal{L} = \|f(x) - \text{SDF}_{\text{GT}}(x)\|_1 + \lambda_{\text{eik}} \|\,\|\nabla_x f(x)\|_2 - 1\,\|_2^2$$ where $\lambda_{\text{eik}} = 5 \times 10^{-3}$. The architecture uses $L = 4$ resolution levels, a per-level feature dimension $d = 2$, and a lightweight ReLU MLP with three hidden layers of width 12. Benefiting from the compact architecture, single-shape SDF training converges in approximately 1 minute on a single NVIDIA V100 GPU.

Key Experimental Results¶

Main Results¶

Evaluations are conducted on three prominent geometric benchmarks: the Stanford 3D Scanning Repository (5 high-precision scans), the ABC dataset (100 CAD models), and Thingi10K (500 complex closed 3D printing models). Geometric accuracy is measured via Chamfer Distance (CD $\times 10^{-6}$) against ground-truth meshes. Self-consistency between the extracted mesh and the neural field is evaluated using Surface-sampled SDF (SSDF $\times 10^{-6}$), Vertex-sampled SDF (VSDF $\times 10^{-6}$), and Angular Difference (AD, degrees) between mesh normals and field gradients.

Table 1: Chamfer Distance (CD $\times 10^{-6}$, lower is better) to ground-truth meshes under Large resolution setting

Method	Stanford	ABC	Thingi10K	Extraction Paradigm
Plain ReLU MLP	5480	4584	3779	Sampling (MC 512)
Analytic Marching (AM)	5475	4570	3775	Analytic (CPWA)
PermutoGrid	3644	3104	2897	Sampling (MC 512)
HashGrid	1659	1854	1763	Sampling (MC 512)
TropicalNeRF	1737	1866	1809	Analytic heuristic
TetraSDF (Ours)	1718	1758	1722	Strict Analytic (CPWA)

Table 2: Self-consistency of meshing methods with the underlying neural SDF (lower is better)

Method	Thingi10K SSDF $\downarrow$	Thingi10K VSDF $\downarrow$	Thingi10K AD (°) $\downarrow$	ABC SSDF $\downarrow$	Stanford SSDF $\downarrow$
MC 256	217.0	98.0	5.44	218.1	289.3
MC 512	65.2	27.0	3.10	66.7	100.1
MC 1024	17.6	1.79	1.68	18.9	20.5
AM	0.020	0.020	0.00	0.021	0.025
TropicalNeRF	144.1	8.7	3.45	136.1	189.0
TetraSDF (Ours)	0.078	0.082	0.00	0.077	0.076

Ablation Study¶

Table 3: Effect of the input preconditioner $A^*$ on reconstruction accuracy (Thingi10K, CD $\times 10^{-6}$)

Configuration	w/o $A^*$	w/ $A^*$	Gain $\Delta$ ($\uparrow$)	Note
HashGrid Baseline	1763	1993	-230	$A^*$ is tailored to tetrahedral geometry and degrades cubic grids
TetraSDF (Ours)	1856	1722	+134	Whitening metric removes directional bias, significantly improving accuracy

Table 4: Mesh complexity vs. accuracy on the Stanford dataset under Small setting

| Extraction Method | Vertices $|\mathcal{V}|$ ($\downarrow$) | Chamfer Distance CD ($\times 10^{-6}$, $\downarrow$) | Reference Benchmark | |-------------------|----------------------------------------|------------------------------------------------------|---------------------| | MC 128 | 24,768 | 1734 | MC 1024 | | MC 256 | 101,337 | 1355 | MC 1024 | | MC 512 | 409,321 | 1290 | MC 1024 | | MT 256 | 396,516 | 1397 | MT 512 | | DC 512 | 676,567 | 1329 | DC 1024 | | TetraSDF (Ours) | 29,557 | 1286 | True zero-level set (compact & artifact-free) |

Key Findings¶

Uncompromised Field Fidelity and Self-consistency: TetraSDF attains SSDF and VSDF values on the order of $10^{-8}$ and an angular difference AD rounding to 0.00°, matching the precision of Analytic Marching while delivering $\sim 10^3\times$ higher self-consistency than MC 1024.
Superior Geometry at a Fraction of Triangle Budget: With only 29,557 vertices, TetraSDF achieves a lower Chamfer Distance (1286) than MC 512 with 409k vertices (1290) and DC 512 with 676k vertices (1329), completely eliminating staircase artifacts.
Geometric Specificity of Metric Whitening: Incorporating $A^*$ provides a +134 CD gain for TetraSDF (notably resolving delicate parts like bunny ears), whereas mistakenly applying it to HashGrid degrades CD by 230, demonstrating that the preconditioner is mathematically coupled to the tetrahedral metric.

Highlights & Insights¶

Reconciling CPWA Theory with Multi-resolution Encoders: By substituting trilinear interpolation with barycentric interpolation on tetrahedra, the work resolves the long-standing dilemma between spectral bias and non-affine interpolation, enabling exact CPWA extraction on high-frequency grid encoders.
Metric Whitening for Unbiased Spatial Learning: The explicit derivation of the spatial metric tensor $M$ and closed-form preconditioning matrix $A^*$ elegantly resolves structural gradient imbalances, providing a general principle for geometric machine learning on irregular spatial tilings.
Efficient GPU Tensorization of Polyhedral Geometry: Grouping shared plane normals and utilizing barycentric masks for perturbation-based neighborhood lookups turns combinatorial polyhedral meshing into highly scalable parallel tensor operations.

Limitations & Future Work¶

Author-Acknowledged Limitations: The current extraction implementation is subjected to minor single-precision floating-point boundary ambiguities at tetrahedral interfaces. Furthermore, the pipeline is evaluated on per-shape SDF fitting rather than integrated end-to-end into multi-view inverse rendering pipelines.
Additional Considerations: Under dense multi-resolution schedules, storing the intermediate candidate edge and vertex sets during edge subdivision incurs substantial transient GPU memory overhead.
Future Directions: Exploring octree-based adaptive tetrahedral subdivisions for unbounded open scenes, and integrating TetraSDF into differentiable inverse rendering to provide exact mesh-based gradients without discrete rasterization proxies.

vs Analytic Marching (AM): AM established analytic CPWA extraction for ReLU MLPs but required massive 10-layer, 90-width networks that suffer from severe spectral bias and struggle with high-frequency details. TetraSDF uses a tiny 3-layer, 12-width MLP over tetrahedral grids, outperforming AM by over $2\times$ in Chamfer Distance (Thingi10K CD 1722 vs 3775).
vs TropicalNeRF: TropicalNeRF applied analytic heuristics to trilinear HashGrids but required local eikonal assumptions and diagonal plane slicing, leading to poor self-consistency (Thingi10K SSDF 144.1 vs TetraSDF 0.078). TetraSDF solves the root cause via affine barycentric interpolation.
vs Neural Marching Cubes / NDC: Differentiable grid variants still rely on discrete voxel grids and surrogate surface parameterizations. TetraSDF extracts the exact continuous mathematical surface encoded by the network parameters.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ Elegant fusion of affine barycentric geometry with CPWA neural network theory.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across three benchmark datasets assessing accuracy, self-consistency, and ablations.
Writing Quality: ⭐⭐⭐⭐⭐ Mathematically rigorous, clear geometric derivations, and well-structured presentation.
Value: ⭐⭐⭐⭐⭐ Provides a foundational milestone for high-precision, discretization-free neural implicit surface meshing.