Gaussian Surfel-Based Representation

• Gaussian surfel-based representation is a method that models 3D surfaces using spatially embedded Gaussian distributions with anisotropic covariances to capture local geometry and uncertainty.
• It leverages multi-resolution mapping, probabilistic fusion, and analytic derivatives to achieve robust real-time SLAM, registration, and high-fidelity surface reconstruction.
• The approach supports applications like LiDAR odometry, scene reconstruction, and differentiable rendering, outperforming voxel-based methods in accuracy and computational efficiency.

A Gaussian surfel-based representation models surface elements (surfels) as spatially-embedded Gaussian distributions in 3D, typically flattened along their local normals or with highly anisotropic covariances to encode surface geometry, appearance, and uncertainty in a unified mathematical form. This approach generalizes classical surface splatting and supports robust, rapid optimization, real-time SLAM, differentiable rendering, and multi-modal data fusion in robotics and computer vision.

1. Mathematical Structure of Gaussian Surfel Representations

A Gaussian surfel is defined by a mean (center) $\mu \in \mathbb{R}^3$, a covariance matrix $\Sigma \in \mathbb{R}^{3 \times 3}$ encoding anisotropic spread (with one axis typically much smaller, representing local surface thickness), a normal vector $n$ (either explicit or inferred), opacity or point weight $\alpha$, and an appearance model, often via spherical harmonics or RGB color $c$.

For example, in surface-oriented contexts, the covariance is factorized as

$\Sigma = R\,\text{diag}(s_u^2, s_v^2, 0)\,R^\top$

where $R$ aligns axes to the local tangent plane $(t_u, t_v)$ and normal $n$, and $s_u, s_v$ specify surfel footprint in-plane (Dai et al., 2024, Strobel et al., 2 Dec 2025, Quenzel et al., 2021).

The density at a point $x$ is given by

$$G(x; \mu, \Sigma) = \exp\left(-\tfrac{1}{2}(x - \mu)^\top \Sigma^{-1}(x - \mu)\right)$$

While classical 3D Gaussians ($\text{rank}(\Sigma) = 3$) model volumetric uncertainty, surfel-based representations enforce or train for $\text{rank}(\Sigma)\approx 2$, producing a mathematically thin, oriented elliptical patch (Dai et al., 2024).

2. Map Organization, Multi-Resolution, and Lookups

Gaussian surfel representations are organized as spatial maps supporting efficient access and dynamic adaptation:

• Hierarchical Multi-Resolution: Maps are stored at $L$ levels of detail, with cell size $m_\ell = m/2^\ell$ for level $\ell$, but constant map volume. Finer levels resolve close or detailed geometry; coarser levels provide global context or fill sparse regions (Quenzel et al., 2021).
• Sparse Data Structures: Efficient spatial lookups are realized via sparse voxel grids (hashed blocks) or permutohedral lattices. The latter offers minimal neighbor sets ($2(d+1)$ in $d$-dimensional space), improving registration performance (Quenzel et al., 2021).
• Adaptive Coarsening: Surfels are dynamically fused or split based on local planarity or degeneracy, as revealed by the eigenvalue spectrum of their covariances. Planar or redundant surfels are coarsened to reduce model size without sacrificing geometric fidelity (Quenzel et al., 2021, Dai et al., 2024).

3. Alignment, Optimization, and SLAM Integration

Surfels support analytic, uncertainty-aware alignment and real-time SLAM:

• Scan-to-Map Registration: Alignment is cast as a Gaussian Mixture Model (GMM), with scan surfels matched probabilistically to model surfels using Mahalanobis distances under the sum covariance,

$$d_{sm}(T) = T \mu_s - \mu_m,\qquad \Sigma_{sm}(T) = \Sigma_m + R \Sigma_s R^\top$$

yielding a likelihood

$$p(e_{sm} | T) = \mathcal{N}(d_{sm}(T); 0, \Sigma_{sm}(T) + \sigma^2 I)$$

(Quenzel et al., 2021).

• Robustness and Soft Association: Registration incorporates outlier modeling, association priors, and geometric compatibility terms (normal/angle/alignment) in the GMM framework, avoiding brittle hard correspondences (Quenzel et al., 2021).
• Continuous-Time Trajectory Estimation: Rigid poses are parameterized as Lie-group B-splines, enabling optimization of complete 6D trajectories over sliding windows of scans, with analytic Jacobians for efficient convergence (Quenzel et al., 2021).
• Key-Frame Fusion and Egocentric Map Management: Surfels from new scans are fused incrementally; the local map is shifted in egocentric coordinates to maintain CPU/memory efficiency during SLAM (Quenzel et al., 2021).

4. Applications and Advantages

Gaussian surfel-based representations are used in:

• Real-Time LiDAR Odometry and Mapping: MARS-Odometry demonstrates sub-frame latency 6D pose tracking by fusing LiDAR scans into Gaussian surfel maps and optimizing trajectory via GMM-based alignment (Quenzel et al., 2021).
• Scene Reconstruction and Surface Estimation: These representations provide closed-form analytic cost functions and derivatives for geometry, supporting high-fidelity, uncertainty-aware surface reconstruction (Dai et al., 2024, Strobel et al., 2 Dec 2025).
• Dense SLAM, Point Cloud Completion, and Multi-Modal Fusion: Gaussian surfels are suitable for multi-resolution mapping, surface completion in ambiguous areas (combining LiDAR and photometric data), and efficient point cloud reconstruction at scale (Strobel et al., 2 Dec 2025, Park et al., 2017).

Key advantages:

• Compact uncertainty-aware modeling of local 3D structure and surface properties.
• Analytic derivatives for differentiable optimization and efficient SLAM.
• Adaptive resolution, robust handling of planar and degenerative geometry, and fast spatial access through advanced lookup structures.

5. Fusion, Uncertainty, and Probabilistic Modeling

Gaussian surfels naturally capture spatial uncertainty and support Bayesian data fusion:

• Measurement Model: For sensor measurements $z$, the likelihood $p(z|x) = \mathcal{N}(z; x, R)$ accommodates anisotropic noise, e.g., high uncertainty along LiDAR beams (Park et al., 2017).
• Fusion via Kalman/Information Filter: The canonical update equations,

$$\Sigma_{\text{post}} = (\Sigma_{\text{prior}}^{-1} + R^{-1})^{-1}, \qquad \mu_{\text{post}} = \Sigma_{\text{post}} (\Sigma_{\text{prior}}^{-1}\mu_{\text{prior}} + R^{-1}z)$$

are employed for each surfel as new points are integrated (Pan et al., 1 Dec 2025, Park et al., 2017).

• Data Association: Candidate surfels for fusion are selected by soft gating on normal and in-plane distances, enabling flexible resolution control without rigid voxelization (Park et al., 2017).

This probabilistic formulation underpins robust, real-time fusion and consistent map growth, with model resolution adapting organically to observed data density and uncertainty.

6. Extensions and Modern Differentiable Pipelines

Recent works extend Gaussian surfel paradigms to more expressive and differentiable graphics pipelines:

• Differentiable Rendering and Deep Integration: Gaussian surfels support efficient analytic rendering, differentiable with respect to geometry and appearance, enabling end-to-end training for view synthesis, inverse rendering, and SLAM (Pan et al., 1 Dec 2025, Xu et al., 2024).
• Spatially-Varying Appearance and Neural Parameterizations: Instead of constant or view-dependent color, surfels may encode local spatial variation with bilinear interpolation, movable kernels, or neural MLPs, substantially improving compactness and rendering fidelity (Xu et al., 2024).
• Enhanced Regularization and Losses: Self-supervised normal-depth consistency and monocular normal priors remedy information loss due to rank-2 local geometry encoding, stabilize optimization, and promote accurate orientation recovery (Dai et al., 2024).
• Multi-modal and Adaptive Fusion: Gaussian surfels facilitate dynamic sensing strategies, e.g., prioritizing ambiguous or low-density regions for scan completion or maintaining balanced representation through density control (Strobel et al., 2 Dec 2025, Quenzel et al., 2021).

7. Benchmarks, Performance, and Comparative Analysis

Empirical studies demonstrate the efficacy of Gaussian surfel-based representations:

• Surface Reconstruction Benchmarks: SurfFill achieves mean Chamfer distances of 0.0055 m and F1-score@5mm of 0.9176 on synthetic data, substantially outperforming 2DGS and 3DGS on fine structure, with low spurious completions (Strobel et al., 2 Dec 2025).
• SLAM and Odometry: The multi-resolution, coarsened surfel maps in MARS-Odometry support real-time continuous trajectory optimization and robust SLAM at high LiDAR rates, with efficient CPU implementations (Quenzel et al., 2021).
• Adaptivity and Robustness: Bayesian fusion and model coarsening contribute to map compactness and error suppression under noisy or ambiguous data, while retaining high geometric fidelity (Park et al., 2017, Quenzel et al., 2021).
• Implementation Efficiency: Sparse data structures and analytic EM optimization strategies keep per-iteration cost low and guarantee scalability to large data volumes and high-rate sensors (Quenzel et al., 2021, Pan et al., 1 Dec 2025).

Benchmarks consistently show superior trade-offs between accuracy, memory efficiency, and compute throughput compared to voxel-based or raw point cloud approaches, validating the fundamental strengths of the Gaussian surfel-based representation.