Gaussian Splatting vs. 3D Mesh

Gaussian splatting vs. 3D mesh is no longer a research debate. It is a processing decision you make on the laptop after the batteries come off the drone, and it changes what the client receives. Both methods start from the same set of overlapping photographs. From there, they diverge completely, in what they store, what they can measure, and what software will open them three years from now.

This guide walks through both pipelines from the ground up. If you have never processed a dataset, the first half explains what each step does and why it exists. If you process weekly, the second half covers capture planning, accuracy expectations, hybrid delivery, and the mistakes that cost reprocessing time.

What happens after the drone lands

Every reconstruction method begins with the same problem: you have several hundred or several thousand images, and the software has to work out where the camera was for each one.

That step is called Structure from Motion, or SfM. The software detects distinctive features in each image, matches them across overlapping frames, and solves for camera positions and orientations simultaneously. The result is a sparse point cloud plus a camera pose for every photograph. Bundle adjustment then refines the whole solution at once, minimising the reprojection error across the dataset.

Georeferencing enters here. RTK or PPK positioning from the aircraft supplies the camera centres, while ground control points constrain the solution to a known coordinate system. Without one of these, your model is internally consistent but floating in an arbitrary frame.

Everything described so far is shared. Gaussian splatting vs 3D mesh only becomes a real fork after the cameras are solved.

How a 3D mesh gets built

The photogrammetry route continues along a path the surveying industry has used for two decades.

Densification. Multi-view stereo takes the solved cameras and computes depth for far more pixels than SfM did. Instead of a few hundred thousand tie points, you now hold tens or hundreds of millions of points, each with a position and a colour.

Surface reconstruction. An algorithm, usually screened Poisson or a Delaunay-based method fits a continuous surface through that cloud. The output is a triangle mesh: vertices connected by edges, forming faces. Crucially, the mesh has an inside and an outside. That distinction is what makes volumes calculable.

Texturing. The software projects the original images back onto the triangles and bakes them into texture atlases. The geometry carries the shape; the texture carries the appearance.

Derived products. Because the mesh is an explicit surface, everything else follows from it. Digital surface models, digital terrain models, orthomosaics, contour lines, cross-sections, and volume reports all come out of the same geometric foundation.

The mesh is therefore a measurement instrument. Its accuracy is documented, its error budget is understood, and surveyors have spent twenty years building acceptance procedures around it.

How Gaussian splatting works

3D Gaussian Splatting was published by Kerbl and colleagues at SIGGRAPH 2023, and it took roughly eighteen months to reach commercial drone software.

Instead of fitting a surface, the method scatters millions of translucent ellipsoids through the scene. Each one, each “splat”, stores a position, a covariance matrix describing its size and rotation, an opacity value, and a set of spherical harmonic coefficients that encode how its colour shifts depending on your viewing angle.

Training is the interesting part. The software renders the current set of splats from a known camera position, compares that render against the actual photograph taken there, and adjusts every parameter to reduce the difference. It repeats this tens of thousands of times. Along the way it splits splats that are carrying too much error and deletes ones contributing nothing.

The result is a model of what the scene looks like from any angle you flew. That distinction explains almost every practical difference below.

Two consequences follow immediately. First, view-dependent colour means reflections, wet asphalt, glass, and polished metal render convincingly rather than smearing. Second, thin objects survive. A guy wire two centimetres across can be represented by a chain of narrow splats, whereas Poisson reconstruction typically either thickens it into a sausage or deletes it entirely.

Gaussian splatting vs 3D mesh: the differences that change your deliverable

The table below is the short version. The subsections that follow explain why each row behaves the way it does.

Accuracy expectations

Photogrammetric meshes have a well-documented error budget. With RTK or PPK positioning and a properly distributed set of ground control points, centimetre-level absolute accuracy is routine on infrastructure sites.

Splat accuracy is a more open question. Because the splats are optimised for visual reconstruction rather than surface fit, geometric error is not the quantity being minimised. Vendor benchmarks published during 2026 have reported mean geometric errors in the high-single-digit centimetre range for building-scale scenes, adequate for visual assessment, insufficient for engineering tolerance work.

Relative geometry is generally sound, since it inherits the camera solution. Absolute surface accuracy is the part still lacking an agreed verification procedure.

How Gaussian splatting vs 3D mesh changes your flight plan

Both pipelines want overlapping imagery. Their preferences diverge in the details, and the flight is where those preferences have to be honoured.

Pix4D’s guidance for splat generation is a useful baseline: aim for 70–80% forward overlap along the flight lines and 60–70% side overlap between adjacent lines. Those numbers will look familiar, because they are close to standard mapping practice. The difference lies in angular coverage.

A mesh mostly needs every surface seen from two or three directions. A splat needs every surface seen from every direction you intend to render it from. Fly a pure nadir grid and the splat will look excellent from above and fall apart at ground level. Add orbits at two or three altitudes, plus obliques at 45 and 60 degrees, and the same scene holds together.

Lighting discipline matters more for splats than for meshes. Since the model learns view-dependent colour, inconsistent exposure teaches it that a surface genuinely changes colour with viewing angle. The visible symptom is floaters: clouds of stray splats hanging in empty space. Therefore lock exposure and white balance, fly in stable light, and avoid the hour when clouds are moving fast.

Motion blur remains fatal for both. Keep shutter speed high, slow the aircraft over detailed structures, and check a sample of images before leaving the site.

Choosing between Gaussian splatting vs 3D mesh by deliverable

Work backwards from what the client actually needs.

The pattern is consistent. Where a number leaves the building and someone acts on it, deliver a mesh. Where a person needs to understand a site without visiting it, deliver a splat.