SurfFill Pipeline for LiDAR Completion

• SurfFill is a LiDAR point cloud completion pipeline that combines LiDAR’s metric accuracy with Gaussian surfel detail recovery to repair ambiguous, under-sampled areas.
• It employs a targeted four-stage process—including ambiguity detection, surfel optimization, sampling, and spatial chunking—to efficiently scale to building-level scenes.
• Quantitative metrics like Chamfer Distance and F1-score reveal SurfFill’s superior ability to recover thin structures and minimize noise compared to traditional methods.

SurfFill is a LiDAR point cloud completion pipeline designed to synthesize high-fidelity point clouds for building-scale scenes by combining the metric accuracy of LiDAR with the detail-reconstruction capabilities of Gaussian surfel splatting. The system leverages a selective, region-focused reconstruction strategy, identifying and repairing only ambiguous, poorly-scanned areas while preserving reliable LiDAR measurements elsewhere. By introducing a divide-and-conquer extension, SurfFill scales to tens of millions of points, outperforming previous shape completion and multi-view photogrammetry methods in terms of geometric precision and thin structure recovery (Strobel et al., 2 Dec 2025).

1. Motivation and Architectural Overview

LiDAR scanning yields exceptional accuracy in flat, texture-free regions but suffers from two central artifact classes: (1) beam divergence leading to mixed pixels and loss of thin structures, and (2) missing returns from dark/reflective surfaces producing spurious holes. Photogrammetric and neural radiance field solutions can recover fine geometry but lack LiDAR’s ground truth metric precision over homogeneous areas. SurfFill addresses these issues through a four-stage pipeline:

1. Preprocessing and ambiguity detection based on point density and image-space uncertainty.
2. Targeted Gaussian surfel-based optimization applied exclusively to ambiguous regions, avoiding redundant reconstruction.
3. Conversion of surfels into discrete point samples, emphasizing densification in incomplete areas.
4. Large-scale processing via spatial chunking and final point set merging.

This architecture allows for accurate, scalable completion of under-sampled structures and large objects without sacrificing LiDAR’s baseline accuracy.

2. LiDAR Artifact Analysis

A fundamental limitation of LiDAR is beam angular divergence. For a vertical half-angle $\theta_v/2$ at range $D$, the laser footprint diameter is given by $d(D)\approx 2D\tan\left(\frac{\theta_v}{2}\right)$ with point density decreasing as $\rho(D)\propto 1/D^2$. When thin features or silhouettes are traversed, multiple surfaces yield time-of-flight collapses, producing depth-averaged “mixed pixels” and loss of edge fidelity. Low reflectance surfaces further cause weak or missing returns, erasing fine details from the scan and generating significant geometric holes or spurious points.

3. Ambiguity Heuristic and Region Selection

SurfFill selectively targets regions of likely error, abstaining from global scene reconstruction. The ambiguity assessment utilizes two criteria:

3.1 Point-Density-Based Ambiguity

Given $\mathcal X\subset\mathbb R^3$ as the input cloud, for each $\mathbf x\in\mathcal X$, $k$ nearest neighbors $\{\mathbf q_i\}_{i=1}^k$ are computed. The inverse-density score

$$p(\mathbf x) = \frac{1}{k\delta}\sum_{i=1}^k\|\mathbf x-\mathbf q_i\|$$

leads to the local density estimate $\rho(\mathbf x) = 1/p(\mathbf x)$. Points for which $$\tau_{\rm min}\leq \rho(\mathbf x)\leq \tau_{\rm max}$$ are marked ambiguous, isolating transitions near holes and thin edges. Thresholds are typically $\tau\approx0.04$.

3.2 Image-Space Uncertainty Map

Monocular normal estimates (Bae et al., ICCV '21) are used to derive per-pixel angular uncertainty

$$U(u,v) = \mathbb E[\angle(\mathbf n(u,v), \mathbf n\text{ in neighborhood})]$$

and binarized by

$$M(u,v) = \begin{cases} 1 & U(u,v)>\sigma \ 0 & \text{otherwise} \end{cases}$$

where ambiguous pixels are permitted aggressive surfel growth ($\sigma\approx0.2$).

4. Gaussian Surfel Reconstruction and Optimization

Ambiguous regions are reconstructed via a set of flat, anisotropic Gaussian surfels, each parameterized by mean position $\boldsymbol\mu_i\in\mathbb R^3$, tangent vectors $\mathbf t_{u,i},\mathbf t_{v,i}$, scale coefficients $s_{u,i},s_{v,i}>0$, and opacity $\alpha_i$. The surfel covariance is

$$\Sigma_i = \mathbf L_i \mathbf L_i^T,\quad \mathbf L_i = [s_{u,i}\mathbf t_{u,i}\;\; s_{v,i}\mathbf t_{v,i}\;\; 0]$$

Optimization proceeds via differentiable alpha-blended compositing into all image views. The overall energy function is:

$$E(\{\mu_i,\Sigma_i,\alpha_i\}) = \sum_v \Big[(0.5+0.5 M_v)(\mathcal L_c + \gamma\mathcal L_g) + (1-M_v)(\alpha \mathcal L_d + \beta \mathcal L_n)\Big] + \rho \mathcal R_s$$

where $\mathcal L_c$ encodes color and D-SSIM loss, $\mathcal L_g$ captures image gradient consistency for thin structures, and $\mathcal R_s$ regularizes surfel size. Constraints include bounds on $s_{u,i},s_{v,i}$, opacity, and adaptive density control for densification/pruning. Inter-splat jitter encourages robust hole-filling.

5. Point Sampling, Densification, and Completion

Discrete points are recovered from optimized surfels by two mechanisms:

5.1 Gaussian Resampling

For each surfel $i$, $m_i$ points are sampled via

$$\mathbf x_{i,j} = \boldsymbol\mu_i + \mathbf L_i\mathbf z_{i,j},\;\;\mathbf z_{i,j}\sim\mathcal N(\mathbf 0,\mathbf I_2)$$

matching local density as needed.

5.2 Between-Surfel Bridging

Inter-surfels are connected by interpolation:

$$\mathbf x' = (1-a)\boldsymbol\mu_i + a\boldsymbol\mu_k,\;\; a\sim\text{Uniform}(0,1)$$

targeting completion of extremely thin gaps.

SurfFill adaptively selects the number of points per surfel to maintain consistency with the original LiDAR density.

6. Divide-and-Conquer Extension for Large Scenes

For scalability to building-level scans, spatial chunking is applied:

• The scene is partitioned into overlapping 3D grid chunks.
• For each chunk, relevant camera and LiDAR data are selected.
• Ambiguity detection, surfel optimization, and sampling are run independently per chunk.
• After processing, non-overlapping sampled points are merged across all chunks, with duplicates within a small radius removed.

This enables processing of scenes containing tens to hundreds of millions of points (e.g., 70M points, 2,000 images over 6 chunks completed in ~75 min on 6×A40 GPUs).

7. Implementation, Evaluation, and Comparative Performance

Initialization occurs with aggressive downsampling of non-ambiguous regions. Surfel sizes are set by

$s_{u,i}=s_{v,i}=\sqrt{p(\mu_i)}\,2^{-1/2}$

and the solver (Adam, learning rate $2\times10^{-4}$, 25,000 iterations/chunk) incorporates periodic densification and no warmup.

Quantitative Metrics

Performance is measured using Chamfer Distance (mean $\sim0.0055$ m on synthetic scenes vs. $0.06$ m for vanilla 2DGS) and F1-score at 5mm tolerance (mean $0.918$ for SurfFill vs. $0.878$ for 2DGS). On specific benchmarks, SurfFill reconstructs thin structures more completely than COLMAP (photo-only) and Neuralangelo (NeRF-only) baselines, introducing minimal excess noise.

Comparative Shape Completion

Deep-learning methods trained on small objects (PointAttN, SnowflakeNet) attain low F1 scores ($\sim0.1$ on a single-chair test). In contrast, SurfFill achieves F1 $\sim0.89$ on the same task.

Summary and Implications

SurfFill leverages the strengths of both LiDAR and photogrammetry by (1) localizing repair regions via ambiguity heuristics, (2) reconstructing these regions with constrained Gaussian surfel modeling, and (3) resampling optimized surfels to generate a completed, high-fidelity point cloud. The method scales robustly to large environments, bridging global metric accuracy and local structure recovery (Strobel et al., 2 Dec 2025). A plausible implication is the adoption of SurfFill in diverse large-scale active reconstruction contexts where scan integrity in the presence of thin structures and under-sampled regions is critical.