Gaussian Splatting and LiDAR: a practitioner's guide to 3D capture

3D Gaussian Splatting (3DGS) is a volumetric rendering technique that represents a scene as a set of 3D ellipsoids (Gaussians), each carrying a position, colour and opacity. Unlike NeRF (Neural Radiance Field), 3DGS does not require a heavy neural network at inference time: rendering is fast, usable in real time on consumer hardware, and scenes can be viewed directly in a WebGL browser.

Published by Kerbl et al. at SIGGRAPH 2023, the method has seen explosive adoption (over 2,000 derivative papers in two years). By 2026, 3DGS is settling into reality-capture pipelines, often as a complement to LiDAR, not a replacement.

3DGS excels at photorealistic rendering and visual immersion. A room-scale capture can be processed in under two hours on a machine with a decent GPU, and viewed in real time. For virtual tours, visual inspection or client communication, the result often surpasses a colourized point cloud.

However, 3DGS does not produce a reliable metric survey. Gaussians are optimised for appearance, not dimensional accuracy. For a submission to a planning authority, a georeferenced LiDAR point cloud remains essential. The winning synergy: use LiDAR for geometric accuracy and 3DGS for the visual envelope — what is called a hybrid LiDAR + GS workflow.

The format ecosystem is evolving rapidly. The glTF standard with the KHR_gaussian_splatting extension (ratification expected Q2 2026 under the Khronos Group) aims to become the universal interchange format. In parallel, OpenUSD (Pixar/Apple/NVIDIA) is developing its own GS support.

Today, the pragmatic formats are:

Jovaxis recommendation: export to PLY today, plan migration to glTF once ratified. For geospatial, favour 3D Tiles + an underlying LAS point cloud.

Capture:

• Xgrids L2 Pro: 32-channel handheld LiDAR scanner (640 K pts/s, ±1–2 cm), direct Gaussian Splatting export via LCC software, with Unity, Unreal and WebGL SDK.
• Trimble X12: survey-grade reference scanner (±1 mm @ 10 m, 2.19 M pts/s), used as an accuracy layer in hybrid workflows.
• LiDAR-equipped smartphones (iPhone Pro, iPad Pro): entry-level for quick captures with apps like Polycam or Scaniverse.

GS Processing:

• Postshot (Jawset): reference GS reconstruction tool, cross-platform.
• Brush (Arthur Brussee, macOS): interactive viewer and editor.
• Luma AI / Polycam / KIRI Engine: accessible cloud pipelines.
• Nerfstudio (gsplat): open-source pipeline for research and automation.

Viewing:

• WebGL viewers (Three.js + gsplat.js, CesiumJS for 3D Tiles).
• Unreal Engine (via plugin) and Unity (via Xgrids SDK).
• NVIDIA Omniverse (emerging USD + GS support).

The recommended approach for a professional survey combines both technologies:

1. Acquire the LiDAR point cloud with a calibrated scanner (terrestrial laser, drone LiDAR, mobile scanner), ensuring georeferencing and quality control.
2. Capture images (from the scanner if camera-equipped, or via supplementary photogrammetry).
3. Generate the Gaussian Splatting from the images (via Postshot, Nerfstudio or Luma AI).
4. Align the GS to the LiDAR point cloud (ICP, common control points, or via Postshot's alignment tool).
5. Deliver a dual-layer viewer: the Gaussian Splatting for immersive navigation, the underlying LiDAR point cloud for accurate measurements — with a magnifier/measurement tool activatable on demand.

This workflow is already used in practice by companies like Cr8ive Media: their 3DGS + point cloud viewer lets users switch between the photorealistic view and the 6.3-million-point LAS cloud for metric measurements.

• Confusing photorealistic rendering with metric accuracy. A GS can look perfect visually yet carry errors of several centimetres. Do not use it alone for a regulatory survey.
• Underestimating GPU requirements. A quality scene needs a recent GPU (RTX 4060 minimum recommended, 12 GB VRAM). Cloud solutions (Luma, KIRI) avoid hardware investment but impose size and privacy limits.
• Neglecting GS–LiDAR alignment. Without common control points or robust ICP, the visual overlay will be offset and measurements distorted.
• Forgetting image rights and GDPR. GS captures the environment with photographic detail. In Europe, data-protection and image-rights rules fully apply — blurring and consent may be required.
• Choosing a dead-end proprietary format. Favour PLY or glTF for long-term data accessibility.

Europe has a favourable ecosystem for deploying Gaussian Splatting coupled with LiDAR:

• Inria (France), the home institution of the original 3DGS paper authors, continues to publish major advances.
• Geo Week 2026 (Denver, February 2026) dedicated several sessions to the impact of GS in geospatial, with strong European participation.
• RIEGL (Austria), Leica Geosystems (Switzerland) and SICK (Germany) are progressively integrating GS-compatible exports into their software ecosystems.
• The Copernicus and INSPIRE initiatives promote geospatial data interoperability; GS could enrich European urban digital twins.

For consultancies and integrators, the 2026–2027 period is one of productive experimentation: GS is no longer a pure research topic, but a differentiating tool for reality-capture deliverables.