SurfFill: Completion of LiDAR Point Clouds via Gaussian Surfel Splatting

TL;DR

SurfFill addresses persistent LiDAR gaps by detecting region ambiguity caused by beam divergence and densifying only those regions with a focused Gaussian surfel approach, guided by RGB-derived uncertainty. The method integrates constrained 2D Gaussian Splatting, sampling from the Gaussian model, and a divide-and-conquer workflow to scale to building-scale scans. Key contributions include an ambiguity heuristic for region selection, a tailored loss and regularization suite for fine-grained reconstruction, and chunked processing for large scenes, all validated against synthetic and real-world data with superior detail restoration. The approach yields accurate completion of thin structures with efficient runtime, outperforming prior results and enabling practical LiDAR data completion at large scales.

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.

Figures (19)

• Figure 1: For a given LiDAR scan, we complete the point cloud by carefully integrating Gaussian surfel-based 3D reconstruction. We analyze erroneous regions and identify ambiguity in the point cloud as well as areas of high uncertainty in captured visual data. Missing structures are reconstructed with a focused 2D Gaussian Splatting technique, followed by filtering and sampling for point cloud completion. We extend this with a divide-and-conquer scheme, allowing us to process this building-scale point cloud in less than 75 minutes.
• Figure 2: High-level functionality of LiDAR and formation of mixed pixels: a) General setup: The sensor emits and receives a light ray and calculates the range via the light travel time, b) Formation of mixed pixels: The laser beam diverges during traveling and thus hits both foreground and background, resulting in a mixed range estimate, c) Raw LiDAR data example with mixed pixels shown in red.
• Figure 3: Our pipeline. We first downsample the LiDAR point cloud and generate uncertainty maps using our ambiguity heuristic. Next, we grow Gaussians into ambiguous regions through our focused surfel splatting technique. Finally, we extract and sample the reconstructed surfels in these regions to generate points, which integrated seamlessly with the LiDAR point cloud.
• Figure 4: Focused Gaussian surfel optimization. Our focused surfel splatting method reconstructs finer details than baseline 2DGS huang20242d.
• Figure 5: Our filtering strategy. We first remove outliers, nearly transparent, large splats or non-densified splats. Next, high-ambiguity Gaussians surfels are removed and the model is filtered based on their distance to the input point cloud. Finally, multiple points are sampled per surfel.