Introduction

In recent years, technologies that convert entire terrains and structures into 3D data using drone aerial photography and laser scanners have been attracting attention in surveying and construction. Advanced 3D modeling, once handled only by specialized surveyors or CG technicians, is becoming more accessible thanks to cloud processing and AI. However, many practitioners still cite concerns such as “expensive equipment and skilled operators are required,” “processing takes a long time even after capture,” and “complex elements like vegetation or water surfaces are difficult to reproduce,” which hinders adoption of the latest technologies in some sites. In practice, many sites face the dilemma of “not being able to quickly 3D-ify the area we need to measure” and “post-processing and sharing of the obtained point cloud data takes time.” As a result, valuable point cloud data often goes underutilized, and information sharing ultimately relies on traditional drawings and photos.

A key to addressing these issues is the emergence of LRTK, which combines smartphones with high-precision positioning, and a new technique for rapidly generating high-fidelity 3D models from images called Gaussian Splatting. LRTK (a positioning system using a smartphone plus a small RTK-GNSS receiver) is an innovative measurement tool in which a pocket-sized device is attached to a smartphone, enabling anyone to easily perform centimeter-level positioning (cm level accuracy (half-inch accuracy)). This makes it possible to acquire data with accurate location information without special equipment or expert skills. On the other hand, Gaussian Splatting (GS), announced in 2023, is the latest 3D generation method that can quickly reproduce photorealistic three-dimensional scenes from multiple photos. This article explains the basic technology and features of Gaussian Splatting in plain terms, explores its advantages through comparison with traditional methods, and introduces its compatibility with data acquired by LRTK and concrete use cases to illustrate the power next-generation 3D modeling brings to field work.

※LRTK: a positioning system using a smartphone + RTK-GNSS receiver (provided by Lefixea Inc.)

Technology and Features of Gaussian Splatting

Gaussian Splatting (GS) is a cutting-edge technique for generating high-quality 3D models from multiple photographs. As the name suggests, it evokes the image of “splatting,” scattering innumerable tiny points (Gaussian distributions) through space and projecting each as a softly blurred colored particle to render surface texture smoothly. By assigning attributes such as color, opacity, and size to each particle, GS can reproduce textures that are indistinguishable from photographs. Unlike traditional point clouds or mesh models, GS directly optimizes the entire scene from images to create the 3D representation, eliminating the need for intermediate point cloud generation or polygonization. As a result, elements that have tended to become noise in conventional methods—such as the fine detail of foliage and reflections on glass or water surfaces—can be rendered as realistic 3D spaces.

Technically, GS draws inspiration from AI-based methods known as NeRF (neural radiance fields), but is characterized by its use of explicit 3D primitives. Whereas NeRF uses large computational resources to learn spatial radiance and density in a black-box manner, GS prepares numerous Gaussians (soft, spherical point primitives) that constitute the scene in advance and adjusts their positions, shapes, and colors to match the photos. Because it does not allocate computational resources to empty regions of space, it is efficient; recent research reports have reduced training (model generation) time to a few hundredths of that of conventional NeRF and achieved rendering at near real-time speeds.

In short, Gaussian Splatting represents a completely different approach from “point clouds that display points as-is” or “meshes that stitch shapes together with surfaces.” It can be seen as a new 3D model representation that combines the lightness of point clouds with the expressive power of meshes. Because many Gaussians overlap to softly cover object surfaces, GS can produce smooth, gap-free models while keeping data volume low. It is also well suited to rendering optimizations on GPUs, and given just the captured images, it has the potential to build and preview models on-site in a short time. These characteristics make GS notable as a next-generation 3D technology capable of pursuing both “photographic realism” and “survey-grade accuracy.”

Comparison with Traditional Techniques and Advantages of Gaussian Splatting

Traditionally, creating 3D models has relied on methods such as “point clouds,” “meshes (polygons),” and “photogrammetry (photo-based composition).” Each is a proven technique, but the following challenges have been pointed out in terms of expressiveness and efficiency.

• Point cloud representation – Point cloud data acquired by laser scanners or photogrammetry is highly accurate, but since it is merely a collection of discrete points, object surfaces can appear sparse. Increasing point density greatly smooths the surface but makes the data volume enormous and difficult to handle. Point clouds themselves have limited color information, so additional processing like projecting photographs or converting to meshes with textures has been necessary.

• Mesh representation – Generating polygonal meshes from point clouds or CAD data covers objects with surfaces and fills gaps, but achieving high detail requires an enormous number of polygons, causing data bloat. Structures with many gaps or transparency, such as tree leaves or chain-link fences, are difficult to reproduce with meshes; trying to represent them can demand massive polygon counts or texture sizes. Moreover, materials that reflect or transmit light, such as water surfaces or mirrors, are hard to represent realistically with static meshes and textures.

• Photo-based composition – Photogrammetry using SfM/MVS methods from drone photos excels at leveraging site photos directly in models. However, its multi-stage processing and heavy computation mean model generation takes time. Manual adjustments such as noise removal and interpolation of missing parts are often required, and operating the specialized software requires skill. Achieving high-precision results also typically involves on-site tasks like placing ground control points (GCPs), which limits responsiveness.

Against these weaknesses of conventional methods, Gaussian Splatting offers many benefits as a next-generation approach. As noted above, GS describes scenes using the position, color, and opacity information of many Gaussian distributions, allowing continuous representation from individual leaves to the sparkle of a water surface. Sparse-looking point cloud displays become filled because each point in GS has a soft spatial extent, producing smooth surfaces without meshes. Additionally, the model-building workflow proceeds seamlessly from image ingestion through optimization, significantly reducing processing time compared to conventional approaches. In some cases, modeling can be completed immediately after on-site capture, enabling additional shooting or on-the-spot data checks as needed.

Regarding data size, GS models—constructed from a collection of lightweight Gaussian primitives—tend to have smaller file sizes than traditional mesh models with equivalent detail. Generated models can be visualized in dedicated viewers, and they can also be converted into point data and exported in conventional formats (LAS/PLY), enabling flexible integration with existing CAD and GIS software. In this way, Gaussian Splatting is expected to prove effective in a variety of field applications as a method that balances expressiveness, efficiency, and data compatibility.

Compatibility between LRTK Data and Gaussian Splatting

To get the most out of GS in the field, high-quality photos and accurate position information are indispensable. This is where data acquired with LRTK proves powerful. Using LRTK, anyone can easily collect photos and point cloud data with position coordinates. For example, if you attach an LRTK receiver to a smartphone and walk around a construction site while photographing, each photo is tagged with highly accurate capture position coordinates at the centimeter level (cm level accuracy (half-inch accuracy)). Traditionally, when creating 3D models with photogrammetry, one had to match common points across multiple photos to estimate camera positions and place target markers or measure known lengths (GCP placement) to set scale. But when photos with known positions from LRTK are used, you can generate a 3D model already aligned to an absolute coordinate system, greatly reducing tedious alignment work.

Furthermore, the ease of acquiring large numbers of photos and point cloud data on-site with LRTK contributes to higher success rates and better quality in GS model generation. Gaussian Splatting performs best when given many overlapping viewpoint images, so an LRTK environment that makes it easy to increase the number of capture angles is ideal. In addition to aerial drone captures, a hybrid data acquisition workflow is possible in which a person equipped with LRTK on the ground walks to photograph fine details, allowing thorough coverage of complex structures’ backsides and forest interiors. Because these data are unified under RTK-referenced coordinates, aerial and ground data can be integrated without misalignment and seamlessly merged into a single model.

Thus, LRTK—which enables anyone to easily collect data with accurate position information—and GS—which can automatically generate high-fidelity 3D models—are truly a next-generation best match. On-site, non-specialist operators can quickly measure and immediately 3D-ify and share data, dramatically shortening the lead time from capture to model review. For example, in disaster response, simply photographing with an LRTK-equipped smartphone can yield a photorealistic 3D model within tens of minutes, enabling accurate spatial information to be shared with remote headquarters. The combination of accuracy, speed, and ease of use offered by LRTK + GS is likely to become the standard for future field measurement and modeling.

Concrete Use Cases

• Construction sites: On infrastructure and building sites, easy on-site surveys using LRTK and the rapid modeling of GS can accurately and realistically record construction progress. For example, 3D-modeling of excavation and embankment as-built conditions overlaid on design data helps detect construction errors early and streamlines earthwork volume calculations. Field measurements that used to require dedicated surveying teams over extended periods can be completed by site technicians using smartphones in a short time, greatly speeding up schedule management and progress reporting. Sharing as-built models via the cloud enables remote offices to monitor construction status, facilitating safety management and client reporting.

• Disaster response: In disasters such as earthquakes or landslides, rapidly 3D-ifying the affected area to grasp the overall situation can be critical for initial recovery planning. Using LRTK-equipped drones to photograph the entire site from above, while workers on the ground photograph around damaged structures with smartphones, allows fast generation of detailed integrated models. GS reproduces rubble piles and collapsed terrain exactly as photographed, enabling remote analysis without entering hazardous areas. Models can be shared on the cloud with municipalities and relief teams to accelerate damage assessment and recovery routing. Moreover, volume measurements of debris and dimensional assessments of damaged areas can be performed accurately in 3D, aiding recovery planning.

• Cultural heritage recording: LRTK + GS is also useful for preserving historic buildings and artworks. Compared with traditional high-skill 3D scanning, precise 3D archives can be created with smartphone photography alone, lowering the barrier to cultural asset surveys. Fine reliefs of sculptures and building ornamentation can be digitally recorded with GS by material, supporting comparative degradation analysis and restoration planning. Full-scale models make it easy to produce replica parts or create VR exhibits, providing a new way to pass valuable heritage to future generations. The reduced need for costly laser scanners and specialist contractors will help promote comprehensive digital preservation of more cultural properties.

• Exterior design: In remodeling houses or gardens, GS-generated as-built 3D models are powerful. Photographing house exteriors and sites with an LRTK-compatible smartphone yields a dimensionally accurate “virtual site” in a short time. Landscape contractors and architects can use this base to examine designs while viewing the actual scenery. Reflective materials such as window glass or water features are reproduced correctly in GS models, making it easier to share completed-image visuals with clients and improving meeting and presentation quality. In real estate, exterior building models can be posted online to provide remote clients with realistic viewing experiences.

• Topographic surveying and environmental analysis: LRTK + GS is effective for broad-area terrain surveys such as mountains and river basins. Areas hidden under trees or in cliff shadows that aerial photos alone cannot capture can be supplemented by ground photography with LRTK devices to comprehensively digitize the terrain. High-resolution models can be used to generate contour lines and longitudinal sections or to simulate earthwork volumes and inundation areas, similarly to traditional point-cloud survey data. Because models are obtained with georeferenced coordinates, they can be overlaid on GIS maps for hazard analysis, or compared across multiple epochs to monitor terrain changes. In disaster prevention, periodic 3D recording of landslide-prone sites or coastal erosion and model comparisons to detect slight terrain changes are promising monitoring applications.

Implementation Flow and Key Points

• Data acquisition: Use GNSS-enabled drones or LRTK smartphones to capture many high-resolution photos of the target area. Planning for roughly 80% overlap between adjacent photos improves 3D reconstruction accuracy. For building exteriors, capture from both ground and air; for terrain, photograph from multiple altitudes to eliminate blind spots.

• Model generation: Load the captured images into dedicated software and the Gaussian Splatting algorithm will automatically generate a 3D model from the photo set. With a PC equipped with a high-performance GPU, processing hundreds of images can complete in on the order of tens of minutes. Because intermediate steps such as point cloud generation and mesh creation are unnecessary, modeling can often be executed with the press of a button.

• Use and sharing: The finished model can be intuitively viewed and measured in dedicated viewers or web browsers. Sharing a URL lets stakeholders inspect the scene remotely. If needed, models can be converted to point cloud data (LAS/PLY) or mesh formats (OBJ, etc.) and imported into conventional CAD/GIS software to overlay on design data, create section drawings, or perform other analyses.

Conclusion

The combination of Gaussian Splatting and LRTK is poised to bring a major transformation in 3D spatial reproduction. High-precision 3D measurement that formerly required specialists can now be realized with familiar tools, and an era in which anyone can handle photorealistic 3D models in a short time is imminent. This approach reduces the cost of outsourcing to surveying firms and complex data processing, directly improving site productivity. By enabling not only complex field measurements but also simple surveys and fixed-point observations with ease, it will dramatically enhance operational efficiency and communication quality.

In fact, in the surveying and construction sectors’ ongoing DX (digital transformation) and digital twin trends, the ease and accuracy of LRTK + GS are highly compatible. Domestically, GS rendering functions are already beginning to be integrated into civil engineering 3D software, and the infrastructure for practical use is steadily developing. Information that was hard to convey with plans and photos alone can be shared three-dimensionally and intuitively using this system. If your site faces challenges in 3D utilization, consider adopting a next-generation workflow that combines LRTK data acquisition and Gaussian Splatting model generation. You will likely experience a level of speed and quality in visualizing 3D spaces that you have not seen before.

In the not-too-distant future, it may become commonplace for anyone on site to scan their surroundings with a smartphone and instantly share 3D models. With LRTK and Gaussian Splatting opening the way, the visualization of field sites will evolve to a new dimension.