Gaussian Splats Explained: The Next Generation of 3D Scene Reconstruction

If you’ve been following spatial data technology in the past 18 months, you’ve probably heard about Gaussian Splatting. The outputs are visually striking — photorealistic, real-time navigable 3D scenes from nothing more than a set of photos or video frames.

Here’s what it is, how it compares to traditional photogrammetry, and what it means for survey work.

What is 3D Gaussian Splatting?

3D Gaussian Splatting (3DGS) is a scene representation technique published by researchers at Inria in 2023. Unlike traditional photogrammetry (which produces point clouds and meshes), Gaussian Splatting represents a scene as a collection of 3D Gaussian functions — essentially small 3D ellipsoids, each with a position, rotation, scale, colour, and opacity.

The key insight: by optimising the properties of millions of these Gaussians to reproduce the input images as accurately as possible, you can create a representation that:

• Renders in real-time at very high quality
• Captures view-dependent effects (reflections, translucency, soft lighting)
• Handles complex scenes that photogrammetry struggles with (reflective surfaces, foliage, fine structures)

Outputs are stored in .splat or .spz files — a relatively new format specific to Gaussian Splatting scenes.

Gaussian Splatting vs. Photogrammetry: what’s the difference?

The key trade-off: Gaussian Splatting produces better-looking scenes, but it’s not a measurement tool. If you need to make accurate measurements from the output (distances, volumes, elevations), photogrammetry is still the right approach.

If your goal is to give someone an immersive visual experience of a space — a heritage site, a construction site progress review, an inspection — Gaussian Splatting is often more compelling.

What Gaussian Splatting is good at

Heritage documentation: Buildings, interiors, cultural sites — Gaussian Splatting captures the visual richness of a space in a way that point clouds can’t match. The photorealistic rendering helps non-technical audiences engage with the documentation.

Construction progress: Monthly captures of a construction site in Gaussian Splatting format give a visually compelling, navigable record of how the site evolved — useful for client presentations and project management.

Inspection delivery: When a visual record of a structure’s condition matters (bridges, facades, industrial equipment), Gaussian Splatting provides a richer visual experience than a point cloud.

Real estate and architecture: Gaussian Splatting of completed or under-construction spaces provides immersive viewing that 360° panoramas can’t match — you can navigate continuously through the space, not just jump between static viewpoints.

What Gaussian Splatting isn’t good at

Geometric accuracy: Gaussian Splatting is not a surveying tool. Don’t try to make measurements from a splat — the geometric fidelity varies across the scene and is not ground-truthed to real-world coordinates.

Large outdoor areas: The technique works best for bounded spaces or structures. Very large open areas (large construction sites, agricultural fields) are better served by drone photogrammetry with GCPs.

Point cloud delivery: If your client needs a LAS/LAZ file for their GIS workflow, Gaussian Splatting doesn’t produce that.

Capture workflow

The capture workflow for Gaussian Splatting is more forgiving than traditional photogrammetry:

1. Video or photos: Dense video walkthrough works well — walk slowly, cover all surfaces, avoid sudden movements. Alternatively, systematic photo capture from many angles.

2. Preprocessing: Extract frames from video (typically 1-2 frames per second), correct lens distortion if needed

3. Camera pose estimation: Run Structure from Motion (using COLMAP or similar) to estimate camera positions — the same first step as photogrammetry

4. Gaussian optimisation: Train the Gaussian scene using a 3DGS implementation (the reference implementation from Inria, or commercial alternatives)

5. Export: Export to .splat or .spz format for browser viewing

Processing typically requires a high-end GPU (RTX 3090 or better). Cloud processing services are available for teams without dedicated GPU hardware.

Viewing Gaussian Splats in a browser

The good news: Gaussian Splatting can be viewed in a browser via WebGL. The .splat and .spz formats are supported by several open-source WebGL renderers.

Platforms like Swyvl support .splat and .spz uploads, rendering them in a browser-based 3DGS viewer. The client gets a link and can navigate the scene in real time in their browser — on laptop or desktop (mobile performance varies, as splat rendering is GPU-intensive).

The file formats: .splat vs .spz

.splat: The original format from the Inria research implementation. Uncompressed — stores each Gaussian’s properties in a fixed-size record. Files are typically 1-5 GB for a typical indoor scene.

.spz: A compressed format developed by Niantic. Typically 2-5x smaller than the equivalent .splat file, with minor quality trade-offs. Increasingly preferred for delivery due to the significant size reduction.

Both formats are supported by current browser-based viewers.

Gaussian Splatting is not going to replace LiDAR or photogrammetry for survey work where geometric accuracy matters. But as a delivery and visualization format — for giving clients an immersive, navigable view of their site — it’s a significant upgrade over point clouds for non-technical audiences.

The technology is developing fast. Expect better geometric accuracy, faster processing, and better large-scale support within the next 12-18 months.