Why "Gaussian Splatting" Now? The Overwhelmingly Real Digital Twin Enabled by LRTK

In recent years, the surveying/construction industry and CG fields have seen growing attention to "digital twins" that faithfully reproduce on-site conditions in virtual space. Using point cloud data captured by drones or laser scanners, or 3D models reconstructed by photogrammetry, projects to remotely inspect sites and use them for construction planning have become widespread. However, traditional 3D models have limits in detail and smoothness when it comes to realistic reproduction; especially with point clouds or polygon meshes, achieving a "being there" level of realism has required advanced processing and very large datasets.

Into this context has emerged a new technique called "Gaussian Splatting". This 3D reconstruction method, originating from recent CG research, draws attention for its ability to render photo-like, realistic 3D scenes quickly by smoothly layering an enormous number of points. The latest cloud services and software are beginning to incorporate it, revealing the possibility that anyone can generate a digital twin with overwhelming realism from drone footage or smartphone measurements.

At the same time, the acquisition of on-site 3D data that forms the basis of those digital twins has also advanced significantly. High-precision 3D measurement that used to require specialized equipment can now be performed easily by site personnel themselves with smartphone-based RTK-GNSS devices known as LRTK. Combining high-precision, geolocated point cloud capture on a smartphone (LRTK) with hyper-realistic visualization via Gaussian Splatting enables digital reproduction and sharing of the on-site "now" at unprecedented speed and accuracy.

This article first explains what Gaussian Splatting is, its technical background, and how it differs from point clouds and meshes in an easy-to-understand way. It then outlines the significance of using this technique in digital twins (site fidelity, lightweight data, smoothness, rendering speed, etc.), discusses compatibility with LRTK and the benefits expected from adoption. We also consider concrete workflows and deployments for generating visual models via Gaussian Splatting from LRTK-acquired point clouds, and discuss future prospects for integration with construction DX, BIM/CIM, cloud sharing, and web visualization. Finally, we summarize how adopting LRTK helps put these cutting-edge technologies to practical use on site.

What is Gaussian Splatting? How it differs from point clouds and meshes

Gaussian Splatting (hereafter GS) is a recently emerged 3D visualization method. Unlike conventional point clouds or meshes, GS assigns to each of the countless points that make up a scene a "softly spreading" Gaussian distribution (a translucent blob with bell-shaped density) and composes the surface by overlapping these blobs. Picture repeatedly layering countless small paint blotches on a canvas to gradually produce a photorealistic image. GS performs this in three-dimensional space, smoothly interpolating between points to reconstruct object shapes. As a result, it can generate extremely smooth, detailed, and photograph-like 3D models.

So what sets GS apart from conventional methods? First, point clouds: point cloud data are collections of coordinates on object surfaces, and while measurement accuracy is high, visualization often retains a grainy look. Gaps between points or flicker (noise) depending on viewing distance and angle can make the model look coarse when moved in real time. Acquiring very dense points improves detail but increases data volume and handling complexity.

Next, meshes (polygon models) form object surfaces by connecting points with triangular faces. Because meshes represent continuous "surfaces," they generally look smooth and are suitable for CAD drawings, BIM model integration, and analyses like volume calculations. However, fine structures (for example, scaffolding pipes or thin wires at a construction site) often get lost or distorted during meshing; hole-filling or simplification can yield shapes that differ from reality. Applying photographic textures to meshes to achieve photorealism requires advanced processing and, depending on model scale, can make the data very heavy and slow to render.

By contrast, Gaussian Splatting takes an intermediate approach: it leverages the position and color information of points like a point cloud but represents each point not as a discrete dot but as a soft surface rather than explicitly creating polygonal faces like a mesh. Each point’s Gaussian "splat" overlaps with others, blurring boundaries so the model appears as a continuous surface. For example, a wall that looked rough in a coarse point cloud appears as a seamless surface in GS because particles blend together. Each splat retains positional accuracy while allowing adjustment of size and shape, enabling faithful reproduction of fine detail. Complex structures that were hard to represent with meshes—scaffolding, piping, tree branches and leaves—can be captured by GS from point-cloud data without loss.

Moreover, GS’s rendering process is highly suited to GPUs. Unlike NeRFs (neural radiance fields) that rely on neural-network-based ray tracing, GS directly projects and composites each splat per view, making rendering extremely fast. Even scenes containing millions of splats can provide smooth, real-time visuals if appropriate LOD (level-of-detail) management is applied. In this way, GS unites the accuracy of point clouds and the smooth appearance of meshes while enabling interactive performance. However, GS-generated models are primarily visualization-focused and not intended for precise CAD editing or geometric analysis (they lack explicit polygon boundaries and surface attributes). With that caveat, GS is positioned as a digital twin representation optimized for "making things look convincingly real."

The value of Gaussian Splatting in digital twins

Using GS for on-site digital twins brings advantages that were previously hard to achieve. Key points include:

• Overwhelming site fidelity: You can obtain vivid, high-resolution 3D models that look as if you are viewing photographs. Because color and material appearance from the real site are reproduced, stakeholders can assess conditions remotely as if they were on site. Fine details such as cracks or sign text are visible on the model, allowing the on-site "as-is" state to be faithfully reproduced in a digital environment.

• Smooth, immersive display: GS models lack point-cloud flicker and do not show polygonal edges like meshes, so moving the viewpoint yields consistently smooth imagery. Navigating inside the model in a 3D viewer is less fatiguing and creates a strong sense of immersion, which is beneficial for VR and AR applications.

• Fast rendering and lightweight data: Thanks to efficient GPU rendering, GS runs responsively. Whereas large point clouds or high-detail meshes traditionally slowed down display, GS can handle large site datasets at near real-time speeds. Because heavy mesh models and high-resolution textures are not required, the overall workflow is lighter and sharing via cloud becomes smoother.

• Faithful representation of complex, fine structures: Thin elements like scaffold tubing, wires, and tree foliage—which often get lost with traditional reconstruction—can be clearly represented with GS. Since GS is point-cloud-based, every measured point can be visualized, so small parts and complex shapes can be preserved as much as possible, improving the completeness of the digital twin for inspection and documentation.

• Intuitive information sharing: Visually realistic digital twins are easy to understand even for non-experts. Conditions that were hard to grasp from point clouds or drawings become immediately clear with a GS model, reducing misunderstandings among stakeholders during construction meetings and smoothing consensus-building. The interactivity—being able to inspect details from necessary viewpoints—offers flexibility that photos or videos cannot, which is another advantage.

Compatibility with LRTK and benefits of adoption

To unlock the full potential of Gaussian Splatting, the acquisition method for the base 3D data is crucial. In that regard, the smartphone surveying solution LRTK is an excellent match for GS. LRTK uses a small RTK-GNSS receiver attached to a smartphone and a dedicated app to position the phone to centimeter-level accuracy during photo capture or LiDAR scanning. Point clouds obtained this way are already in accurate absolute coordinates (world coordinates) from the start, and distortions during scanning are corrected in real time. As a result, the point clouds and photos produced align with the actual geodetic system without post-processing, and the high-precision positioning is preserved directly in GS-generated models. Photogrammetry-derived 3D models typically end up in arbitrary local coordinate systems, but using LRTK yields digital twins aligned to the site coordinate system from the outset.

Measurement with smartphone + LRTK is extremely convenient and agile as a data capture method for GS. Without preparing specialized laser scanners or elaborate capture rigs, site personnel can collect high-precision point clouds and photos simply by walking the area with a smartphone. Images captured will carry RTK-derived geotags, enabling fast and stable results when generating point clouds or GS models later via photogrammetry. Georeferenced data ensures correct scale for lengths and areas, so the resulting GS models are more than just images—they can be used as measurable digital twins. Point clouds scanned over multiple days will align precisely, making it possible to stitch large GS model areas together without positional offsets.

LRTK also has high AR affinity, enabling placement of acquired 3D data over the real world. For example, overlaying an LRTK-acquired and GS-processed as-is model on a tablet’s AR app allows immediate review of past site conditions in situ. Similarly, preemptively GS-modeling existing structures and displaying them alongside planned BIM models in AR makes intuitive old-vs-new comparisons possible. It is precisely because LRTK provides high-precision alignment that GS’s photoreal models can be used effectively for such spatial overlays. LRTK already includes AR features for guiding pile-driving and for overlaying design data on collected point clouds for as-built checks; combining those with GS’s smooth as-built visuals yields even clearer, more persuasive site visualizations.

In this way, LRTK is a key platform for applying cutting-edge techniques like Gaussian Splatting on site. Adopting it enables 3D measurement that balances "sufficient accuracy and overwhelming convenience," dramatically improving the quality and usability of resulting digital twins. The ability to reduce initial costs compared to dedicated equipment and the ease of use by non-specialists are major attractions. Workflows such as uploading scanned data to the cloud to generate and share GS models immediately become feasible. With LRTK, quick scans during routine construction management or inspections can be turned into digital twins and shared with stakeholders for instant review and discussion. This directly supports the rapid information sharing and decision-making that construction DX aims to achieve.

Future outlook for construction DX: BIM/CIM integration and cloud sharing

Digital twins enabled by the combination of Gaussian Splatting and LRTK are expected to play a central role in promoting construction DX going forward. In particular, integration with BIM/CIM could realize workflows that handle planning and actual conditions seamlessly. For example, overlaying an LRTK-acquired GS model captured during construction with the BIM under design would quickly reveal deviations or construction errors. While point cloud data are already sometimes imported into BIM software for clash detection, GS models—because of their rich visual information and intuitiveness—should become even more useful references across design, construction, and maintenance phases. In the future, technologies may mature to efficiently extract CAD drawings from GS point-cloud models or to refine GS rendering using BIM models, further narrowing the gap between reality and design data.

On the cloud sharing and web display front, GS broadens the applicability of digital twins. Hosting high-fidelity 3D models in the cloud and allowing stakeholders to interactively change viewpoints via a web browser would enable remote site inspection during online meetings. Large 3D datasets that previously required specialized viewers or high-performance PCs are easier to stream with GS’s efficient compression and rendering techniques, making mobile device viewing feasible. Soon, site supervisors might routinely check high-detail models on a tablet and share status and instructions on the spot. Overseas, initiatives are already appearing to construct city-scale digital twins using GS and publish them on web platforms. In CG domains like entertainment and VR training, photoreal 3D spaces will generate new experiential value. In infrastructure inspection, disaster prevention, and urban planning, the value of sharing and analyzing realistic 3D spatial information via the cloud will continue to rise.

These technology trends align with government-led digital reforms. In Japan, i-Construction and CIM introduction guidelines recommend 3D use on site, and in the future photo-like point-cloud models may be expected as standard deliverables. Real-time sharing of digital twins among stakeholders to support fast decision-making could contribute to productivity improvements and work-style reforms in the construction industry. Gaussian Splatting is a key technology that can accelerate this realization. Combined with 5G and cloud computing advances, a world where remote stakeholders can instantly grasp site details is within reach. In the coming years, digital twins using Gaussian Splatting could become the industry standard and an indispensable tool for driving DX.

Conclusion

Technologies for on-site digital twins are evolving rapidly. Among them, Gaussian Splatting delivers immediately perceptible, overwhelming realism and elevates the value of digital twins to the next level. The role of LRTK as the practical foundation for using these capabilities on site is also significant. A time is approaching when anyone can perform high-precision site scans with a smartphone and generate smooth, beautiful 3D models from that data.

"Why Gaussian Splatting now?" The answer is that technological maturity and increasing on-site needs intersect—now is an ideal time to adopt it. GS combines expressive power and speed that previous methods could not achieve, and early adaptation can improve operational efficiency and provide differentiation from competitors.

LRTK is a practical solution for taking that first step. It brings advanced technology to the field simply and makes cutting-edge methods like GS usable at the site level. Why not take this opportunity to start experiencing overwhelmingly realistic digital twins through LRTK? State-of-the-art realism should bring new discoveries and value to your site. As the boundary between reality and digital fades, let's pioneer the future site visions that this innovative technology will create together.