The New Standard for Construction DX: Smart Sites Enabled by Gaussian Splatting × LRTK

In the construction industry, digital transformation (DX) has accelerated in recent years, and the shift from traditional 2D data like drawings and photos to visualization through 3D data is gaining momentum. Capturing and sharing sites in 3D makes it easier to convey the finished image intuitively, discover design errors, and align understanding among stakeholders. As a result, 3D utilization—leading to improvements in quality, safety, and efficiency—is becoming a foundational technology for construction DX.

Among these trends, a technology drawing attention is “Gaussian Splatting.” Meanwhile, a solution called “LRTK” has emerged that enables easy measurement of high-precision point clouds using a smartphone, and it is spreading rapidly. This article explains what Gaussian Splatting is in an accessible way and unpacks why it is gaining attention in construction DX now. It also describes the characteristics of point cloud data obtainable with LRTK, the benefits of combining Gaussian Splatting and LRTK, concrete on-site use cases (construction management, consensus building, inspections, etc.), and the appeal of visualization accuracy and lightweight rendering with this technology, along with future prospects. Finally, we explain why you should start acquiring spatial information with LRTK now from a forward-looking perspective. The tone is aimed at non-technical readers as well, so please use this as a reference for building future smart sites.

What is Gaussian Splatting?

Gaussian Splatting (also abbreviated as 3DGS) is an advanced 3D representation technique that reconstructs a three-dimensional scene from multiple photos and can rapidly generate novel viewpoint images. Its distinctive feature is that it does not convert a scene’s volumetric information into polygons or other surfaces; instead, it projects and composites countless elements represented by Gaussian distributions directly into space for rendering. In other words, it can be considered a type of point cloud method where each point is treated as a softly spread “3D Gaussian.”

Specifically, feature points are first detected from multiple input-view images to create a point cloud. For each point, surrounding color and brightness information are blurred and averaged using a Gaussian function, and colored elliptical points are placed in 3D space as if splattering paint onto a canvas. By layering tens of thousands of these “Gaussian-blurred points,” a smooth, realistic 3D space is reproduced that looks as if it were photographed. Unlike polygonal models that produce faceted surfaces, the transitions between points naturally blend, enabling photorealistic representation that is pleasing to the human eye.

What makes Gaussian Splatting revolutionary is its reconstruction accuracy and processing speed. It can capture textures that were difficult to express with conventional 3D CG—such as the transparency of water, reflections on glass, and metallic sheen—as 3D data. It can also render extremely fine details like overlapping leaves or individual animal hairs. Moreover, these high-fidelity models can be generated and displayed in relatively short timeframes, which is another major advantage. Research reports indicate that from about one minute of footage, roughly 30 minutes of machine-learning processing produced a high-quality 3D model. Recently, simple apps capable of completing capture-to-model generation on smartphones have appeared, enabling results to be checked on-site with 1–2 minutes of capture and several minutes to tens of minutes of computation.

This technology is part of the lineage of AI-based image-to-3D reconstruction methods that began with NeRF (Neural Radiance Fields) around 2020. While NeRF was groundbreaking, it required long training times that made real-time use difficult. Gaussian Splatting represents scenes explicitly with Gaussian-type point clouds without using neural networks, achieving both drastically faster processing and high rendering quality. The real-time rendering approach published in 2023 made a significant impact on industry, accelerating applied research and demonstrations across many fields, including construction.

Why is it being noticed in construction DX now?

There are two main reasons Gaussian Splatting is attracting attention in construction DX: the increasing necessity for digital technology adoption and the practical-level evolution of 3D technology.

From an industry-wide perspective, the Ministry of Land, Infrastructure, Transport and Tourism (MLIT) began mandating BIM/CIM (use of 3D models) for some public works from fiscal 2023, making 3D data utilization effectively essential. In addition, labor shortages and the aging of skilled workers have made efficiency improvements and know-how transfer via digitalization and automation urgent. Relying on paper drawings and 2D photos makes it increasingly difficult to complete complex projects on time and budget, so productivity gains through DX are unavoidable. Utilizing 3D site data has been shown to enable consistent efficiency from planning to construction, inspection, and maintenance. For example, using point cloud measurement data of as-built conditions for component dimension checks reportedly reduced work time and costs by about 73% compared to traditional manual measurement. Automatically color-coding differences on scanned 3D data can reveal millimeter-level deviations that the human eye would miss, greatly contributing to quality assurance and reduction of rework. In this way, 3D visualization has become an indispensable element in the productivity revolution of construction sites.

On the technical side, until recently, recording a site in 3D required expensive laser scanners or specialized contractors. But today, anyone can acquire 3D point clouds with smartphones or drones. For example, combining LiDAR sensors or high-performance cameras in modern smartphones with RTK-GNSS receivers allows high-precision point cloud collection simply by walking the site. LRTK consolidates these technologies into a solution that enables absolute-coordinate point clouds to be obtained with one hand and without surveying expertise. Moreover, AI techniques like Gaussian Splatting now make it realistic to automatically generate detailed 3D models from captured photo sets. Tasks that once took hours in photogrammetry software can now be completed quickly with Gaussian Splatting, and the resulting models are lightweight and easy to handle. In short, the barriers to acquiring and using 3D data have dramatically lowered, and starting site DX now offers a chance to maximize benefits from the latest technologies and fundamentally streamline workflows.

Characteristics of point clouds and spatial information obtainable with LRTK

LRTK is a solution that uses a high-precision GNSS (RTK) receiver attached to a smartphone to enable anyone to easily measure three-dimensional point cloud data. Point cloud data represent objects and terrain as a collection of numerous points in 3D, where each point includes position coordinates (X, Y, Z) and sometimes color information (RGB) or return intensity. Like pixels forming an image, the more numerous (denser) the points, the more finely shapes can be reproduced. LRTK scans the surroundings with a smartphone’s LiDAR sensor or camera while using RTK-GNSS to determine the measurer’s position with centimeter-level accuracy, thereby producing point clouds with absolute coordinates (latitude, longitude, elevation).

This “point cloud scan × RTK positioning” yields data with several important characteristics. Because every point has geographic coordinates (e.g., a global geodetic system), the acquired point cloud can be directly overlaid onto maps or design coordinate systems. Traditionally, point clouds from laser scanners or smartphone LiDAR were highly accurate in relative shape within a site but required target placement or post-processing alignment to match public coordinate systems. With LRTK, you can obtain 3D data tied to global coordinates from the start, greatly simplifying the reconciliation of site surveys with drawings or BIM data. For example, merging point clouds collected on different days is straightforward because each dataset shares a common coordinate frame, enabling the easy creation of integrated 3D models for wide-area terrain or large structures. RTK positioning also provides very high accuracy—typically within a few centimeters—so measurements of distances, areas, and volumes on the resulting point cloud model can achieve accuracy comparable to physical measurements. In other words, LRTK generates real-scale, high-precision digital spatial information by fusing 3D scanning and positioning.

LRTK point clouds can also include color information from photographs. By applying images taken with a smartphone camera as textures to the point cloud, you obtain rich 3D data that reflect the site’s actual colors rather than a monochrome set of points. Such color point clouds and 3D mesh models auto-generated from point clouds can be easily viewed and shared in the cloud without specialized software. LRTK lets you upload measurement data to a dedicated cloud, share it with stakeholders via a browser-based 3D viewer, and immediately measure distances, cross-sections, and volumes. This end-to-end support—from site capture to office verification and instantaneous quantity calculations for ordering or construction management—makes LRTK a strong enabler of everyday use of point cloud data.

Gaussian Splatting × LRTK: What happens when they combine?

So, what happens when Gaussian Splatting and LRTK-derived 3D point clouds are combined? In short, they deliver a “smart site” that digitally reproduces the real space almost exactly. The fusion of Gaussian Splatting’s photorealistic rendering and LRTK point clouds’ high-precision positioning produces a digital twin that balances precision and visual clarity.

Historically, there has been a trade-off: measurement-oriented point clouds were accurate in shape but visually coarse, while photo-based 3D models were visually realistic but lacked reliable dimensional accuracy. Combining the two yields data that leverage the strengths of both. LRTK point clouds provide the skeleton (precise geometry), while Gaussian Splatting offers the facial expression (detailed color and texture). For example, you can take precise measurements from any viewpoint using the point cloud and simultaneously inspect the same area with Gaussian Splatting’s high-fidelity visuals to read fine concrete cracks or labels on equipment. Photo textures supplement details that are hard to discern from point clouds alone, enabling reproduction on data that closely matches what is seen on-site.

Put another way, you obtain a new class of site data that unites accurate measurements with immersive visuals. This combination—delivering high-level measurement, recording, and visualization together—could become the new standard for construction DX.

On-site use cases and future outlook

The smart site data produced by Gaussian Splatting and LRTK can be valuable across many construction workflows. Below are examples in construction management, consensus building, and inspection & maintenance, with their expected benefits.

• Construction management: Recording and sharing progress with 3D models streamlines verification of as-built conditions and quality control. For instance, scanning newly installed structures with LRTK and comparing the resulting point cloud + GS model to the design BIM model can instantly reveal shape deviations and construction mistakes. Tasks that once required comparing cross-sections and site photos can be performed intuitively on a digital twin, reducing rework. Remote managers can also review site models from the office and issue instructions easily. Visualizing daily progress in 3D helps close the information gap between site and office and supports quicker decision-making.

• Consensus building: 3D is also effective for communication with clients and local residents. For example, visualizing the pre-construction environment realistically with a GS-enhanced point cloud and compositing the planned building into it communicates the finished scale intuitively. Where drawings or CG perspective images failed to convey scale, a 3D model integrated with the real surroundings is far more persuasive. Stakeholders can discuss while taking virtual site tours, making it easier to obtain agreement on design changes or during public briefings. In the future, this could be overlaid on-site with AR glasses so people can experience the completed project in place.

• Inspection & maintenance: GS×LRTK data are powerful for infrastructure and building maintenance. Regular 3D scans allow you to accumulate and compare aging changes digitally. For example, you can quantify whether bridge cracks have widened year over year or how equipment inside tunnels has altered by analyzing differences between models. Because point clouds include positional information, you can accurately locate anomalies on the actual structure and inform repair planning. Inspections in high or confined spaces can be conducted safely without entering them by using drone×LRTK GS models. Looking ahead, AI-based automatic crack detection or degradation prediction using multi-year scan datasets are promising applications that widen the field further.

These examples illustrate how realistic 3D site data from Gaussian Splatting and LRTK can add new value across construction projects. As visualization accuracy and shareability increase, all stakeholders can discuss and decide based on a common “virtual reality,” ultimately improving project-wide productivity and transparency.

The appeal of Gaussian Splatting’s visualization accuracy and low-load rendering

Another reason Gaussian Splatting stands out is its ability to combine high visualization accuracy with low rendering load. As noted before, the technique layers photographic information in a point-like form, and because each point has a soft spread, the overall realism is not severely degraded even if the point count is reduced. Traditionally, producing high-detail 3D models required enormous numbers of polygons or points, making data size and rendering load huge. With Gaussian Splatting, each point can represent information over a certain area, enabling a smooth appearance with relatively modest data sizes. By carefully tuning color, opacity, and shape per point and layering them, near-photographic imagery can be rendered in real time.

This lightness is a major advantage for field data use. If high-detail textured 3D models can be displayed in a web browser rather than a specialized viewer, stakeholders can inspect site data on their PCs or tablets without installing software. Gaussian Splatting suits such use cases well, and browser-based GS viewers have emerged in recent years. With just a URL click, 3D site models can be accessed, greatly smoothing internal and external information sharing. On the distribution side, an optimized compression format for Gaussian Splatting (SPZ format) has been proposed to maintain high quality with lightweight files. These trends suggest that Gaussian Splatting not only achieves realism but also realizes easy-to-handle 3D data, which is highly attractive.

Low rendering load also expands the potential for real-time applications. Reports already show real-time display at over 30 FPS, and it is conceivable that one day we could scan a site and display a 3D model instantly. For example, if GS models could be generated and cloud-shared while walking a construction site, as-built checks and remote support could be performed on the spot. The combination of high visual quality and speed offered by Gaussian Splatting will likely make it increasingly practical in daily workflows.

Future R&D and expected application fields

The potential unlocked by combining Gaussian Splatting and LRTK will continue to broaden. From a research perspective, advances are expected in capturing dynamic scenes and enhancing real-time processing. Today the primary use is static site reproduction, but in the future it may be possible to record and replay construction progress as a time-series 4D model (4D Gaussian Splatting), or to capture heavy machinery and worker movements in real time to aid safety management. Another promising application is smart inspection, where high-fidelity data are analyzed by AI to automatically detect changes or anomalies. While image-based AI for crack and rust detection has progressed, combining it with GS would allow precise localization in 3D space and quantitative evaluation of degradation over time. As R&D advances, the accuracy of automatic modeling and automated analysis should improve dramatically, further accelerating construction DX.

Applications will extend beyond construction. Gaussian Splatting is already being demonstrated in general contractors, shipbuilding, infrastructure maintenance, cultural heritage archiving, and education. For cultural heritage, the ability to quickly produce high-resolution 3DGS models from photos is invaluable for digital preservation or VR exhibits. In urban planning, rapid generation of district-scale 3D models from drone imagery could be used for sunlight simulations or evacuation planning. In automated construction machinery and robotic work contexts, realistic digitalization of site environments is essential; GS×LRTK-derived high-precision site models could be used to train AI robots for accurate task execution. In entertainment, the technique is promising for incorporating real-world scenery into games and film. Because it enables whole-world scanning and virtual experiences, it aligns well with metaverse and VR/AR content, expanding commercial opportunities.

Thus, the fusion of Gaussian Splatting and precise spatial measurement is poised to generate innovative solutions across many fields. For those involved in construction DX, staying abreast of these technological trends and adopting them early in business process reforms will be increasingly important.

Closing: Why you should start acquiring spatial information with LRTK now

The smart site future envisioned by Gaussian Splatting and LRTK is not distant science fiction but is already within reach. You might wonder, “Isn’t it still too early for my company to adopt this?” However, as past technological revolutions show, the earlier you ride the DX wave, the greater the advantage. If you begin acquiring and accumulating high-precision spatial data now, you will gain a significant edge when these technologies become industry standards.

First, data are a new asset. Accumulating 3D point clouds and models of your sites establishes a foundation for future analyses and applications. If you later want to view past site conditions in 3D but have no prior data, it will be too late. LRTK enables easy digital archiving of current site conditions, which becomes a valuable resource for future renovations or inspections.

Second, knowledge of site DX cannot be accumulated instantaneously. Workflows for 3D scanning and data utilization must be optimized through on-site practice. Starting LRTK point cloud measurements now lets your organization build know-how so that when the technology matures, you will be ahead in practical application. Conversely, late adoption risks being left without expertise while others take 3D utilization for granted.

Third, the immediate benefits are substantial. LRTK-based simple point cloud scans are an investment in the future and already deliver measurable returns today, such as more efficient surveying and drawing preparation and enhanced site records. As noted earlier, using point clouds for as-built management alone can substantially reduce time and improve quality. Thus, adopting now is not merely enduring for the future but a practical choice because it’s useful today, creating a virtuous cycle that also prepares you for tomorrow.

Fortunately, user-friendly tools like LRTK have made high-precision surveying and point cloud technology—once the domain of specialists—available to everyone on-site. The ease of starting with just a smartphone is a major help for DX beginners. Taking even a small first step is crucial. To prepare for the new standard of digital twinization of sites, consider starting spatial information acquisition with LRTK now. The journey to future smart sites has already begun.