Practical Geo Scan guidance on Gaussian Splatting for Surveying for architects, developers and project teams.
Gaussian splatting has moved rapidly from computer-graphics research into practical reality-capture workflows.
The reason is easy to understand once you see a good result.
A complex building, façade, street scene or plant space can be reconstructed as a highly realistic three-dimensional environment that is fluid to navigate and often feels much closer to being on site than a conventional point cloud.
For architects, engineers, heritage teams and property professionals, that has obvious appeal.
But surveying introduces a different question:
Does a convincing visual reconstruction also provide dependable geometry?
Not necessarily.
That distinction is central to how Geo Scan is approaching Gaussian splatting.
A reconstruction can look extraordinarily real without every part of it carrying the measurement confidence required of a survey. Visual fidelity and geometric accuracy are related questions, but they are not the same question.
3D Gaussian Splatting is a method of representing a scene using a large number of three-dimensional Gaussian primitives.
Rather than describing the environment as a conventional polygon mesh or simply displaying discrete XYZ points, the method optimises many small, oriented volumetric elements with properties such as position, scale, orientation, opacity and appearance.
When those elements are rendered together from a viewpoint, they reproduce the appearance of the captured scene.
The original 3DGS research was developed for novel-view synthesis: generating high-quality views of a scene from camera positions that were not part of the original image set.
That origin matters.
Gaussian splatting was not invented as a surveying instrument.
Its primary objective was compelling, efficient scene representation and rendering.
For built-environment work, that creates an interesting opportunity — but it also explains why visual quality alone should not be used as evidence of survey accuracy.
In a typical image-led workflow, a scene is photographed from many overlapping viewpoints.
Those images are used to estimate camera positions and establish the spatial relationships needed to optimise the Gaussian representation.
Capture may use:
Good coverage still matters.
The reconstruction can only infer the scene from the information made available to it. Areas that are hidden, poorly photographed, reflective, transparent, moving or visually ambiguous can remain difficult.
That should sound familiar to anyone working with photogrammetry.
The capture processes are related, but the resulting scene representations and their intended strengths are different.
Traditional photogrammetry uses overlapping photographs to reconstruct three-dimensional geometry, commonly producing a point cloud, mesh, orthophoto or textured surface.
Gaussian splatting also uses image relationships, but its representation is optimised particularly strongly for rendering the appearance of a scene from different viewpoints.
This can make a splat look remarkably natural.
Fine visual detail, foliage, irregular objects and complex surfaces can sometimes be represented in a way that feels more convincing than a conventional textured mesh.
But a better-looking representation is not automatically a better measurement surface.
It depends on the output.
If the requirement is an orthophoto, metrically controlled surface model, conventional point cloud or established downstream photogrammetric product, photogrammetry has a mature survey workflow.
If the requirement is immersive visual understanding and rapid novel-view rendering, Gaussian splatting can be extremely compelling.
For some projects, the strongest answer may be to derive more than one output from a well-planned image capture, rather than treating the technologies as competitors.
Geo Scan's existing heritage photogrammetry guidance explains why image-led capture is already valuable where texture, material and visual detail matter.
Terrestrial laser scanning directly measures ranges to surfaces from known scanner positions.
A survey network can then be registered, controlled and checked to produce a dense point cloud suitable for measured survey work.
Gaussian splatting is fundamentally different.
Its appearance is reconstructed from imagery and optimised for rendering. Depending on the workflow, there may be useful geometry associated with the reconstruction, but that does not make the rendered splat equivalent to a TLS point cloud.
There is no useful universal yes-or-no answer.
Recent research is actively investigating metric and geospatial accuracy for Gaussian-splatting reconstructions, including building and UAV datasets. The results show that geometric evaluation is possible, but they also reinforce an important surveying principle: accuracy has to be demonstrated for the workflow and output being used, not assumed from visual quality.
For Geo Scan, that means we would not describe a Gaussian splat as survey-grade simply because it aligns well on screen.
Where dimensions matter, the dataset needs an appropriate scale, coordinate framework, control and independent checks.
Our guide to terrestrial laser scanning accuracy explains the same principle for conventional scanning: the completed survey matters more than a headline technology specification.
Software may allow a user to click points and display a distance.
That does not, by itself, establish the accuracy of the measurement.
There are several separate questions:
A measurement tool is therefore not the same thing as a measurement specification.
This is one reason Geo Scan sees splatting as particularly valuable alongside authoritative survey data.
Survey deliverables are often excellent at answering precise technical questions.
A CAD drawing can describe plans, elevations and sections clearly. A BIM model can organise interpreted building information. A point cloud preserves dense three-dimensional observations.
But those outputs do not always give a remote project team the immediate feeling of the actual space.
A high-quality splat can potentially fill that gap.
A user may be able to move through a realistic representation and understand relationships that would take longer to interpret from drawings alone.
Image-led capture records visual information that can be valuable for design teams, heritage work, coordination and revisiting site conditions.
A useful navigable record can reduce the need to rely on memory, isolated photographs or repeated site visits for questions that are primarily visual.
Plant spaces, historic interiors, irregular façades and cluttered areas can be difficult to communicate with conventional photographs because the viewer loses the relationship between images.
A spatially coherent visual representation can make those photographs much more useful.
Geo Scan is exploring Gaussian splatting as part of a broader controlled reality-capture workflow, rather than as a replacement for established measured-survey methods.
The approach we are interested in has two complementary layers.
Where the project requires dependable measurement, the principal geometry is established using appropriate survey methods — typically static TLS, total-station control and GNSS where relevant.
That gives us the coordinate framework against which other information can be assessed.
A separate image capture can be optimised for the qualities Gaussian splatting does particularly well: coverage, colour, visual continuity and intuitive navigation.
Surveyed targets or other identifiable control can be included where they add value.
The objective is not to claim that control magically turns the splat into a laser scan.
It is to create stronger relationships between the visual reconstruction and the survey framework, and to test what accuracy is actually achievable.
The opportunity is not to choose between a survey and a visually rich reconstruction. It is to understand when the project benefits from having both.
Targets with known survey coordinates can provide useful reference information within an image-led workflow.
Depending on the processing route, they may help with:
But control needs to be used carefully.
No — not by themselves.
A handful of targets may demonstrate agreement at those locations without proving the quality of every reconstructed surface between them.
This is the same reason a registration residual should not be treated as the sole proof of a laser-scanning survey.
Control is evidence.
It is not a substitute for understanding the behaviour of the complete dataset.
This is where the technology becomes particularly interesting for Geo Scan.
Imagine a project in which the building is captured with a survey-controlled TLS network.
That point cloud is used for the measured deliverables and remains the authoritative source for geometry.
During the same attendance, a structured image capture is collected for Gaussian splatting.
The two datasets can then serve different purposes:
This is not duplication if each dataset answers a different project need.
It is a layered record of the existing building.
Our LiDAR scanning services guide explores a related hybrid principle: use faster supplementary capture where it adds value, while retaining an authoritative survey framework for the information that needs it.
Plant rooms are an obvious research use case.
They contain dense pipework, ducts, equipment, valves, labels and services that can be difficult to understand from a conventional plan.
TLS is extremely effective for capturing their geometry, but the point cloud may not communicate every colour, label or material relationship as naturally as high-quality imagery.
A splat could provide a complementary visual environment through which a designer can revisit the space remotely, while the point cloud or BIM model remains available for geometry.
Our Design Museum MEP void survey demonstrates the kind of congested services environment where comprehensive reality capture can be particularly valuable.
This also leads directly into Geo Scan's forthcoming dedicated Plant Room & MEP Laser Scanning service page.
Heritage environments are another strong candidate.
The value of a historic interior is not only its geometry.
Colour, texture, decoration, material changes, irregular surfaces and condition can all help a project team understand the asset.
Photogrammetry already plays an important role here, particularly for textured geometry and orthographic outputs.
Gaussian splatting may add another layer: a fluid visual representation that makes a complex historic space easier to revisit and communicate.
The limitation remains the same.
Where conservation decisions depend on dimensions or deformation, the visual reconstruction should not quietly replace controlled measurement.
Image-led capture can also be attractive for façades, courtyards and larger external scenes.
The viewer gains a more natural sense of the building in context than a single elevation or isolated photographs can provide.
However, external work introduces its own challenges:
For dimensional façade work, Geo Scan would still choose the measurement methodology around the required deliverable rather than around the visual appeal of the reconstruction.
Potentially — but that needs careful wording.
Image capture can be extremely rapid compared with deliberately measuring every feature individually.
That can make Gaussian splatting attractive as a way of preserving broad visual context while the survey team is already on site.
It may also reduce later visits for questions that can be answered visually from the captured record.
But there is no value in saving site time if the project later discovers that information needed for measurement was never captured to the required standard.
The efficiency comes from matching the capture method to the information requirement.
That is the same principle behind our approach to mobile LiDAR.
A realistic workflow needs to acknowledge the limitations.
If an area cannot be seen adequately in the source imagery, the reconstruction cannot be expected to contain dependable information about it.
Glass, mirrors, polished metal and other view-dependent surfaces can be difficult because their appearance changes with viewpoint.
Image-based camera estimation benefits from distinctive visual information. Large plain walls and repetitive environments can make this harder.
People, vehicles, foliage and other moving objects can introduce inconsistencies between images.
Strong exposure differences, moving shadows and changing illumination can affect visual consistency.
A reconstruction can be visually excellent while still requiring separate work to establish scale, control and geometric reliability.
Point clouds, DWG, IFC and other established survey formats have mature downstream workflows. Gaussian-splat formats and viewers are still evolving quickly, so long-term accessibility and client delivery need consideration.
A splat is also not an interpreted drawing or model.
It can show what the environment looked like at the time of capture, but it does not automatically identify:
Those are interpretation tasks.
For projects requiring conventional deliverables, the captured survey information still needs to be translated into outputs such as 2D CAD or Scan to BIM.
Gaussian splatting can make the source information easier to understand.
It does not remove the need to interpret it.
The most interesting deliverable may not be a splat in isolation.
A future project package could combine:
That would allow different members of the project team to use the representation best suited to their task.
An architect checking a dimension may use the survey drawing or point cloud.
A designer trying to remember how services pass through a congested room may open the visual viewer.
A heritage consultant may use high-resolution imagery and photogrammetric outputs to inspect material detail.
The value is in connecting those resources rather than pretending they are interchangeable.
Because the technology is developing quickly, Geo Scan's R&D needs to be empirical.
Useful tests include:
The aim is to find the boundary between impressive visualisation and dependable project information.
That boundary may move as the technology improves.
It is already relevant to professional reality capture.
That does not mean every Gaussian-splat workflow is ready to replace established survey deliverables.
Research published since the original 2023 3DGS paper is increasingly examining geometric, metric and geospatial performance. That is encouraging because it moves the conversation beyond appearance alone.
For professional use, Geo Scan would still ask the same questions we apply to any survey technology:
Technology changes.
Those questions remain.
Gaussian splatting sits within a growing family of ways to record existing environments.
Terrestrial laser scanning is strong where dense, dependable geometry is required.
Mobile LiDAR can offer fast supplementary spatial capture in suitable controlled workflows.
Photogrammetry can provide metric image-derived products, textured geometry and orthographic imagery.
Gaussian splatting can provide exceptionally rich visual scene representation and navigation.
CAD and BIM turn captured information into interpreted project deliverables.
The opportunity for surveying practices is not to declare one of these technologies the winner.
It is to understand how they can be combined intelligently.
Innovation in surveying is not about replacing proven methods for the sake of something new. It is about knowing where new methods can make the overall survey more complete, useful and efficient.
The original 3D Gaussian Splatting method was introduced by Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler and George Drettakis in 3D Gaussian Splatting for Real-Time Radiance Field Rendering, ACM Transactions on Graphics, 2023.
More recent academic work has begun evaluating 3DGS specifically for metric reconstruction, building documentation, UAV mapping and cultural heritage. This is an active research area, so individual results should be read in the context of their capture method, control, dataset and validation procedure rather than treated as universal performance specifications.
Gaussian splatting is one of the most interesting developments in image-led reality capture.
Its ability to create realistic, navigable representations of complex real-world environments has clear potential for architecture, engineering, heritage and building documentation.
For surveying, the important step is to separate what looks convincing from what has been demonstrated to be geometrically reliable.
Geo Scan's approach is therefore deliberately hybrid.
Use established survey methods where dependable measurement is required. Use image-led techniques where they improve visual understanding, coverage or efficiency. Introduce control and independent checks where appropriate. Then connect the outputs so that the project team can use each dataset for the task it does best.
The right data, captured at the right level of accuracy, as efficiently as the project allows.