Local GP Signed Distance Field (L-GPDF)

• L-GPDF is a probabilistic 3D representation that infers the signed distance from any point to a surface using a mixture of locally trained Gaussian Processes.
• It employs kernel selection and sparse Gaussian Process techniques to achieve accurate, scalable shape reconstruction and real-time mapping.
• Integration with volumetric frameworks like OpenVDB enables rapid voxel lookup and dynamic online fusion with rigorous uncertainty quantification.

A Local Gaussian Process Signed Distance Field (L-GPDF) is a probabilistic, continuous, and memory-efficient representation of 3D geometry, in which the signed distance to the surface at any point in space is inferred as the output of a mixture of locally trained Gaussian Processes (GPs). In contrast to conventional explicit surface representations (e.g., meshes or point clouds), L-GPDFs permit smooth signed-distance field (SDF) queries at arbitrary locations, encode surface uncertainty, and naturally support sparse or streaming updates. This framework has become central to state-of-the-art object modeling, scene reconstruction, and real-time robotic mapping, by enabling high-fidelity geometric inference with rigorous uncertainty quantification, while mitigating the cubic scaling bottlenecks of global GP models.

1. Mathematical Foundations of Local GP Signed Distance Fields

An L-GPDF models the surface geometry by covering the domain of interest with $K$ anchor points (“local centers” or “reference points”) $\{\mathbf C_k\}_{k=1}^K$. Around each anchor, a local directional distance field is defined via a zero-mean Gaussian process on the sphere. Specifically, for an anchor $\mathbf C \in \mathbb R^3$, the distance $d$ from $\mathbf C$ to the surface in direction $\boldsymbol\psi=(\phi,\theta)$ is modeled as: $$f(\boldsymbol\psi) \sim \mathcal{GP}(0,\,k(\boldsymbol\psi,\boldsymbol\psi'))$$ with $k$ a positive-definite covariance kernel. For each observed surface point $\mathbf P_i$, compute its direction and radial distance from the anchor as

$$\boldsymbol\psi_i = (\phi_i, \theta_i),\quad d_i = \|\mathbf P_i - \mathbf C\|$$

and collect the training pairs $\{\mathbf C_k\}_{k=1}^K$0. The likelihood is Gaussian: $\{\mathbf C_k\}_{k=1}^K$1 For a query direction $\{\mathbf C_k\}_{k=1}^K$2, the GP predictive posterior mean and variance are

$\{\mathbf C_k\}_{k=1}^K$3

where $\{\mathbf C_k\}_{k=1}^K$4 and $\{\mathbf C_k\}_{k=1}^K$5. The mixture field over $\{\mathbf C_k\}_{k=1}^K$6 is given by

$\{\mathbf C_k\}_{k=1}^K$7

with $\{\mathbf C_k\}_{k=1}^K$8 (spherical coordinates), and mixture weights $\{\mathbf C_k\}_{k=1}^K$9 depending on proximity or local covariance.

Signed distance values are obtained by applying a sign convention: for points inside the anchor hull, the sign is negative; otherwise, positive. This convention provides a meaningful SDF across the object, with smooth transitions due to weight overlap (Sapoutzoglou et al., 1 Apr 2026).

2. Kernel Selection, Hyperparameter Learning, and Gaussian Process Mixture Structure

The expressivity and multiscale sensitivity of L-GPDFs are governed by the choice of covariance kernel $\mathbf C \in \mathbb R^3$0. Common choices include:

• Squared-Exponential (RBF):

$\mathbf C \in \mathbb R^3$1

• Rational-Quadratic (RQ):

$\mathbf C \in \mathbb R^3$2

• Matérn, periodic, polynomial kernels have also been evaluated.

All kernel hyperparameters ($\mathbf C \in \mathbb R^3$3, $\mathbf C \in \mathbb R^3$4, $\mathbf C \in \mathbb R^3$5, $\mathbf C \in \mathbb R^3$6) are estimated by maximizing the log-marginal likelihood: $\mathbf C \in \mathbb R^3$7 Empirical results demonstrate that the RQ kernel outperforms alternatives for shape reconstruction, particularly in capturing both sharp and smooth geometric features (Sapoutzoglou et al., 1 Apr 2026).

Anchor placement is pivotal. Clustering (e.g., k-means) provides an initial reference set, but semantically informed methods (skeletonization, topology decomposition) yield higher fidelity in objects with complex morphology.

3. Algorithmic Structure: Inference, Partitioning, and Acceleration

Inference at a query point $\mathbf C \in \mathbb R^3$8 proceeds by evaluating each local GP at the appropriate direction, weighting the predictions, and applying the sign rule. The per-query cost for a naive implementation is $\mathbf C \in \mathbb R^3$9 ($d$0 anchors, $d$1 points per anchor). To address the cubic $d$2 training scaling per GP, several strategies are deployed:

• Sparse GPs: Utilize $d$3 inducing points, reducing complexity to $d$4 per local GP.
• Kernel Interpolation & Subspace Methods: Alternatives such as KISS-GP or random Fourier features accelerate training and querying by approximating or factorizing kernel matrices.

Overall, the mixture structure—combining many small, local GPs—facilitates parallelism and enables both tractable training and real-time querying, unlike global GPs whose inference cost is prohibitive for large $d$5 (Sapoutzoglou et al., 1 Apr 2026, Wu et al., 2024).

4. Integration with Volumetric and Streaming Frameworks

A key practical advance is the embedding of L-GPDF into high-performance volumetric data structures, notably OpenVDB. In this paradigm, the volumetric workspace is partitioned into small, fixed-size clusters (leaf nodes), each storing its own local GP trained on voxel centers near the surface. This architectural pairing yields:

• Constant-Time Voxel Lookup: The shallow, fixed-depth B+ tree structure of OpenVDB allows $d$6 access to the relevant local GP for any spatial query.
• Real-Time Online Fusion: Each observation frame triggers a round of local L-GPDF training (microseconds per block), raycasting, and probabilistic fusion of new SDF estimates into the global VDB. The fusion step employs variance-weighted averaging, integrating measurement uncertainty directly into the SDF.
• Marching Cubes and Global GP: After fusion, zero-crossing surface points are extracted and a “Global GPDF” (also local per leaf) is updated, enabling continuous ESDF queries even in partially observed or unobserved regions (Wu et al., 2024).

This design provides both rapid, scalable access to local SDF predictions and global consistency, supporting large-scale mapping, dynamic environments, and arbitrarily sized scenes.

5. Empirical Performance, Applications, and Comparative Evaluation

L-GPDF methodologies achieve state-of-the-art results on standard shape and mapping benchmarks:

Benchmark	Chamfer Distance $d$7	F-score	Reference
ShapeNet: Planes	$d$8	$d$9	(Sapoutzoglou et al., 1 Apr 2026)
ShapeNet: Chairs	$\mathbf C$0	$\mathbf C$1	(Sapoutzoglou et al., 1 Apr 2026)
ShapeNet: Sofas	$\mathbf C$2	$\mathbf C$3	(Sapoutzoglou et al., 1 Apr 2026)
Cow & Lady (RGB-D)	$\mathbf C$4 mm	–	(Wu et al., 2024)

L-GPDF outperforms TSDF and Euclidean Distance Transform (EDT)-based methods on reconstruction accuracy (lower Chamfer and RMSE) and provides continuous, differentiable uncertainty quantification. For instance, on Cow & Lady and Newer College LiDAR datasets, VDB-GPDF achieves 2.5 mm Chamfer and 0.015 m RMSE in 2D slices, surpassing TSDF and EDT (Wu et al., 2024).

Foundational applications include:

• 3D object shape representation from sparse point clouds (Sapoutzoglou et al., 1 Apr 2026)
• Online scene mapping for robotics and navigation (Wu et al., 2024)
• Any scenario requiring dense, uncertainty-aware geometry processing

6. Relation to Global GPDFs and Other SDF Representations

Global GPDF methods, in which a single large GP models the SDF over the entire domain, incur substantial cubic time and memory penalties, as well as practical restrictions on dynamic scene updates (Choi et al., 2023). Practitioners have shown that covering the workspace with overlapping cells or clusters and training independent local GPs allows real-time performance and streaming data integration, properties not achieved by earlier global-only designs.

Conventional SDFs based on TSDFs or EDTs are limited by projective artifacts, lack of uncertainty, and often only approximate true Euclidean distance. By contrast, L-GPDFs provide fully probabilistic, continuous SDFs with explicit uncertainty and improved geometric fidelity, while maintaining practical scalability (Wu et al., 2024).

7. Limitations and Prospective Directions

The accuracy of L-GPDFs depends on the density and placement of anchor points; a plateau is typically reached at 16–32 anchors for objects of moderate complexity, with little gain from further increases (Sapoutzoglou et al., 1 Apr 2026). The current bottleneck remains the per-leaf cubic scaling, though sparse GP and kernel-interpolation methods can mitigate this.

A plausible implication is that further gains may be realized by adaptively varying anchor/block density based on local geometric complexity, or by hierarchical GP mixture models. Integration with high-dimensional data streams (e.g., multi-modal sensory fusion with tactile or appearance attributes) and coupling with information-theoretic planning for active exploration represent further extensions, as pursued in global GPDF frameworks (Choi et al., 7 Jul 2025), though formal L-GPDF variants specialized to these tasks remain in development.

L-GPDF stands as a foundational framework in object-centric 3D geometry representation, reconciling statistical rigor, uncertainty quantification, and scalability for modern robotic and computer vision applications.