SurfFill: Completion of LiDAR Point Clouds via Gaussian Surfel Splatting (2512.03010v1)

Abstract: LiDAR-captured point clouds are often considered the gold standard in active 3D reconstruction. While their accuracy is exceptional in flat regions, the capturing is susceptible to miss small geometric structures and may fail with dark, absorbent materials. Alternatively, capturing multiple photos of the scene and applying 3D photogrammetry can infer these details as they often represent feature-rich regions. However, the accuracy of LiDAR for featureless regions is rarely reached. Therefore, we suggest combining the strengths of LiDAR and camera-based capture by introducing SurfFill: a Gaussian surfel-based LiDAR completion scheme. We analyze LiDAR capturings and attribute LiDAR beam divergence as a main factor for artifacts, manifesting mostly at thin structures and edges. We use this insight to introduce an ambiguity heuristic for completed scans by evaluating the change in density in the point cloud. This allows us to identify points close to missed areas, which we can then use to grow additional points from to complete the scan. For this point growing, we constrain Gaussian surfel reconstruction [Huang et al. 2024] to focus optimization and densification on these ambiguous areas. Finally, Gaussian primitives of the reconstruction in ambiguous areas are extracted and sampled for points to complete the point cloud. To address the challenges of large-scale reconstruction, we extend this pipeline with a divide-and-conquer scheme for building-sized point cloud completion. We evaluate on the task of LiDAR point cloud completion of synthetic and real-world scenes and find that our method outperforms previous reconstruction methods.

• The paper introduces an ambiguity heuristic to detect missing regions and guide focused surfel splatting for improved LiDAR point cloud completion.
• It integrates Gaussian surfel splatting with uncertainty mapping to accurately reconstruct fine geometric details while reducing artifacts.
• Empirical results show significant gains in F1-score and Chamfer Distance compared to existing methods on both synthetic and real-world datasets.

SurfFill: Completion of LiDAR Point Clouds via Gaussian Surfel Splatting

Introduction and Motivation

LiDAR-based point cloud acquisition is a de facto standard for metric 3D reconstruction, providing high-precision data particularly in flat, feature-sparse regions. However, LiDAR suffers from systemic artifacts: missed small geometric structures, ambiguous depth readings due to beam divergence, and high rates of incompleteness for dark or reflective materials. These incomplete regions are problematic in applications requiring full geometric fidelity. Photogrammetric methods, such as multi-view stereo (MVS) and neural radiance field techniques, offer detailed inference for feature-rich areas but lag in global precision. The SurfFill pipeline addresses this complementarity by integrating LiDAR's structural precision with focused appearance-driven surfel splatting methods, completing missing structures and ambiguous regions.

Figure 1: High-level functionality of LiDAR and formation of mixed pixels, illustrating beam divergence and subsequent mixed range estimates.

Technical Contributions

Ambiguity Heuristic for Region Detection

A central foundation of SurfFill is an ambiguity heuristic, which identifies transition regions at the borders of missing structures in sparse/semi-dense LiDAR point clouds. For each point, the ambiguity score $p(\mathbf{x})$ is computed via inverse local density using $k$-nearest neighbors, allowing robust localization of ambiguous boundaries without requiring ground-truth labels. This enables the detection of regions requiring geometric completion, overcoming the limitations of direct structural inference.

Focused Gaussian Surfel Splatting

Building on recent advances in 2D Gaussian Splatting (2DGS) [huang20242d], SurfFill constrains the surfel-based densification and optimization to the ambiguous regions prioritized by the ambiguity heuristic. The system employs several adaptations:

• Maximum screen-space gradients for surfel cloning yield improved detail recovery in sparse photometric datasets.
• Integration of monocular surface normal uncertainty maps, computed via predicted normal angular errors [bae2021estimating], further restricts optimization to image-space uncertain regions identified as incomplete or ambiguous.
• Custom loss formulation, featuring enhanced edge reconstruction ($\mathcal{L}_g$), scale regularization for fine structure fidelity, and masked depth/normal regularization to avoid oversmoothing in border regions.
• Probabilistic sampling of training views, up-weighting those providing visibility of ambiguous regions, enables focused GPU resource utilization.

  Figure 2: Pipeline overview showing ambiguity-based downsampling, uncertainty mapping, focused surfel splatting, and final sampling for point cloud completion.

  

  Figure 3: Focused surfel optimization reconstructs finer details in ambiguous regions compared to conventional 2DGS.

Filtering and Sampling

Post-optimization, SurfFill employs a hierarchical filtering protocol:

• Removal of outlier surfels and non-critical splats using scale, opacity, and ambiguity thresholds.
• Spatial filtering based on geometric proximity to input LiDAR data, ensuring only genuinely missing structures are filled.
• Multimodal surfel sampling—including intra-surfels and interpolated samples—to densely bridge gaps while avoiding redundancy with existing LiDAR points.

  Figure 4: Filtering strategy applied post-splatting, pruning outliers, non-densified surfels, and sampling for completion.

Large-scale Scene Completion

The system generalizes to building-scale scenes via a robust chunking ("divide-and-conquer") approach inspired by [kerbl2024hierarchical, lin2024vastgaussian], partitioning scenes into axis-aligned bounding boxes. Each chunk is independently optimized with its relevant images and points, enabling parallelization across multiple GPUs (demonstrated at city-scale datasets with hundreds of millions of points).

Figure 5: Visualization of chunk selection and data partitioning for large scene processing.

Empirical Results

SurfFill demonstrates strong numerical superiority relative to leading baseline techniques. On synthetic and real-world scenes, SurfFill attains an average F1-score of 0.9176, outperforming 2DGS (0.8781), 3DGS (0.8795), GOF (0.8731), COLMAP (0.8141), and Neuralangelo (0.7765) (2512.03010). Chamfer Distance metrics similarly show improved geometric accuracy. Qualitative evaluations reveal superior recovery of fine structures (legs, fence bars), with minimal extraneous noise.

Figure 6: Comparative results showing SurfFill's recovery of missed small structures across synthetic test scenes.

Figure 7: Performance on real-world datasets, highlighting SurfFill's capability to reconstruct missed elements such as chair legs and beams.

Ablation studies further validate the importance of each pipeline component (e.g., filtering yields an average F1-score boost of 2.5%, maximum gradient densification +1.6%).

A direct comparison with state-of-the-art point cloud completion networks (SnowflakeNet, PointAttN) on extracted chair/table objects highlights dramatic generalization failures for these models (F1: 0.1054 vs. SurfFill 0.8918), underscoring the limitations of learning-based completion for complex, high-density LiDAR scenes.

Figure 8: Direct point cloud completion methods (SnowflakeNet, PointAttN) fail against SurfFill on complex structures.

Scalability and Runtime

SurfFill is designed for practical deployment in large-scale settings, capable of completing office buildings and 350m bridge scenes within 2–3 hours on standard A40 GPUs. Chunk overlap and point extension mitigate chunk boundary artifacts, albeit with slight degradation in excessive subdivision, suggesting an area for future improvement.

Figure 9: Large office scene processed in 6 chunks, demonstrating scalable completion with integration of new points (highlighted in green).

(Figure 10, Figure 11)

Figure 10: Massive 350m bridge dataset partitioned in 18 chunks; Figure 11: Visualization of added points (green), indicating integration even in challenging low-quality scenarios.

Limitations and Future Directions

SurfFill inherits several limitations from photometric reconstruction: dependence on image quality and coverage, susceptibility to pose noise, and inability to address dynamic elements or severe lighting variability. Current pose robustness is competitive, but improvements are possible via adaptive pose correction or incremental scene buildup. The method's efficacy is attenuated in extremely sparse or visually challenged datasets, as highlighted in bridge-scale experiments.

Figure 12: Evaluation of pose noise robustness across methods.

Theoretical and Practical Implications

SurfFill's ambiguity-driven, focused surfel splatting paradigm demonstrates that hybrid multimodal approaches leveraging both LiDAR and RGB data, coupled with adaptive region selection, can achieve detailed, scalable, and high-fidelity completion unattainable by existing point cloud upsampling or photometric-only methods. The methodology establishes a baseline for practical deployment in industrial-scale surveying, robotics, BIM, and AR contexts where completeness of metric scenes is critical.

The theoretical innovation—a surface ambiguity heuristic for targeted reconstruction—opens future research avenues in generalized artifact detection for multimodal sensor fusion, adaptive scene chunking for distributed optimization, and uncertainty-aware loss formulation for improved boundary recovery.

Conclusion

SurfFill presents an advanced, ambiguity-guided LiDAR completion pipeline integrating focused Gaussian surfel splatting, uncertainty mapping, and robust hierarchical optimization, setting state-of-the-art accuracy for point cloud recovery in synthetic and real-world scenarios. Its scalable design positions it for immediate adoption in industrial and academic scene reconstruction workflows, while its methodological advances suggest promising future developments in multimodal geometric completion and artifact-aware optimization.