Analytical photogrammetry, which involves analyzing photographs based on the principles of perspective and projective geometry, can be traced back to the early 19th century, when photography first emerged. Building upon these foundational principles, modern 3D scanning methods are generally classified into four categories: laser scanning, structured light, LiDAR, and photogrammetry.

Among these methods, the laser, structured light, and LiDAR scanning approaches operate by detecting the distance traveled by reflected light or laser beams and the angles of projection, enabling rapid and precise measurements of an object’s detailed geometry and absolute dimensions. However, because only surfaces that are directly accessible to light or laser beams can be captured, scanning large areas requires multiple passes followed by the alignment of individual scans. This is further compounded by the high cost of professional 3D scanners, often costing tens of thousands of dollars, and the requirement for dedicated post-processing software, which limits accessibility. To address this limitation, LiDAR components were miniaturized and commercialized in 2020 as sensors integrated into Apple iPhones and iPad devices, thereby enabling mobile-based 3D scanning. While these mobile versions also exhibit certain drawbacks, including reduced accuracy compared to professional scanners, shorter detection ranges, and susceptibility to environmental factors like sunlight and weather, they remain highly accessible, offering rapid 3D scanning capabilities using only a mobile device (Cho and Hwang, 2024).

In parallel, photogrammetry, a photographic measurement technique, reconstructs 3D models by applying optical principles and projective geometry techniques to overlapping images captured from multiple angles around the subject. From these images, 3D coordinates and color values are calculated based on both the distance between the camera and object and the spatial distribution of points within each image. However, because the resulting 3D measurements are derived from relative differences between photographs, accuracy diminishes in occluded regions and in areas with fine geometric detail. Moreover, generating high-density point cloud data from numerous high-resolution images entails substantial processing time (Cho and Hwang, 2024).

Building on this foundation, numerous researchers worldwide have investigated the application of 3D scanning in digital forensics and crime scene reconstruction. Esposito et al. (2023) analyzed case studies utilizing laser scanning (LS) technology for forensic investigation. Their work demonstrated the utility of LS in creating immediate scale plans, reconstructing cold cases, clarifying suicidal intent through simulation, and studying COVID-19 transmission pathways. They concluded that LS is a fundamental tool for ‘crystallizing’ the investigative scene, proving its high value in forensic pathology.

Concurrently, studies focusing on accessible mobile technologies have also been conducted. Sheshtar et al. (2025) compared the results of mobile-phone-based photogrammetry and LiDAR scanning when reconstructing indoor crime scenes. They scanned a mock indoor crime scene, comprising dummies, furniture, reference markers, and bloodstains, using an Apple iPhone 15 Pro Max with the Recon-3D application for mobile LiDAR scanning and the Pix4D solution for photogrammetry. Scans were performed under varying lighting conditions (daytime and nighttime) and scan durations (5 and 10 min) as independent variables. The results demonstrated that both scanning methods achieved accuracy levels sufficient for forensic use. Specifically, the 5-min daytime LiDAR scan yielded the highest accuracy. However, performance limitations were observed under challenging conditions; nighttime photogrammetry deteriorated when using flash, and LiDAR scans required approximately 9 h of cloud processing, thereby underscoring the critical influence of environmental conditions and processing time.

Further, Yang et al. (2024) investigated the use of mobile photogrammetry for 3D wound analysis by scanning clay blocks, pig skin, and bone models using the Scaniverse 3D Scanner app on an iPhone. They also assessed the accuracy and practical applicability of the resulting depth data. All models were successfully reconstructed in under 1 min on average, and the measurements no statistically significant differences from physical measurements, indicating that mobile photogrammetry could serve as a valuable supplementary tool for documenting wounds at crime scenes.

Based on this evidence, several crime scene units and research institutions in South Korea, including the National Police Agency and the National Forensic Service, have begun using 3D scanners to document real crime scenes during forensic casework (Seoul Public News, 2016). Furthermore, in July 2025, the police announced the launch of Policelab 3.0, a next-generation initiative aimed at developing advanced scene analysis technologies integrating 3D scanning and AI from 2025 to 2030 (MBN, 2025). However, despite this institutional momentum, the high cost of 3D scanner equipment and stringent security protocols have limited widespread deployment across all forensic investigation sites.

Recent advances in AI technologies have led to the emergence of approaches such as NeRF and 3DGS, driving substantial shifts across the 3D scanning and graphics industries. Among these, NeRF, introduced by Mildenhall et al. (2020), represents a deep neural network that predicts the required volume density (σ) and color values for rendering 3D scenes by applying deep learning to 2D images. It processes photographs and videos captured with smartphones or ordinary cameras and generates 3D information far more rapidly than conventional photogrammetry owing to its deep-learning-based training. Instead of constructing a 3D model, it directly renders 2D images or videos from 3D representations. Moreover, by incorporating optimizations proposed in follow-up investigations, it reduces training times to enable image rendering within 30 min to 1 h. This technique has gained further traction with the advent of 3DGS. For instance, Kerbl et al. (2023) presented a volume rendering method, that directly visualizes 3D volumetric data using Gaussians, thereby circumventing the need for mesh conversion. Moreover, because 3DGS produces high-resolution images from low-density 3D data without relying on dense point clouds or large mesh files, it enables real-time high-definition rendering and demonstrates broad application potential in areas such as virtual studios, virtual reality and extended reality platforms, animation, video content production, and gaming.

In a previous study, Fukuda et al. (2024) quantitatively compared the strengths and limitations of NeRF and photogrammetry, noting that while conventional photogrammetry performs well for objects with clearly defined patterns or texture-rich surfaces, NeRF yields superior reconstructions for objects with monochromatic, low-texture surfaces, as well as metallic, reflective, or translucent materials.

Traditional photogrammetry requires extensive processing time to generate high-density point cloud data and presents challenges in managing massive datasets. Similarly, Neural Radiance Fields (NeRF) are limited by slow training speeds. In contrast, 3D Gaussian Splatting (3DGS) enables rapid training and real-time rendering, making it highly suitable for forensic investigations where securing the ‘critical time window’ is essential.

While international studies attempting to apply NeRF and 3DGS technologies to crime scene reconstruction and digital forensics are beginning to emerge, research in this specific domain remains relatively scarce.

For instance, Malik et al. (2024) evaluated the feasibility of adopting NeRF as an alternative to photogrammetry during forensic autopsy documentation by comparing reconstructions produced by Instant NeRF and photogrammetry across diverse forensic subjects, including discolored corpses, metal containers, vehicles, and simulated crime scenes. Their results revealed that although the NeRF-based model exhibited lower overall spatial resolution, it achieved higher fidelity in reconstructing forms, colors, and textures of varied surface materials, such as reflective metal surfaces, discolored skin, and plastic components.

Extending the scope of NeRF applications, Remondino et al. (2023) reviewed its use in crime scene video analysis, using the CCTV-based UCF Crime Dataset to compare diverse NeRF variants and advanced models focused on multi-object composition, deformation synthesis, and lighting effects.

Further, to evaluate the efficiency, safety, and innovative potential of 3D reconstruction techniques in capturing crime scenes, Rangelov et al. (2024) performed 3D reconstructions of simulated laboratory scenes, mock crime scenes at the Dutch Police Academy, and fire scenes in collaboration with the Dutch Twente regional fire department using photogrammetry and NeRF. These reconstructions exhibited high resolution, confirming that both techniques offer adequate fidelity for future crime scene investigations.

Overall, the above studies indicate that 3DGS can rapidly reconstruct real-world scenes and objects in high resolution using only images and photographs. In forensic contexts, where photography plays a crucial role in preserving scene information, this capability is particularly relevant as multi-angle photos and videos captured onsite can later be used to reconstruct crime scenes using 3DGS and support investigations at any stage. Moreover, because 3DGS produces compact high-quality data, managing and preserving forensic evidence becomes more efficient. Additionally, given that an increasing number of 3D software platforms now support 3DGS data and that these data can be converted into meshes and traditional point clouds, the applicability of 3DGS is expected to expand across branches of forensic research that rely on general-purpose 3D software.

The 3DGS technique, as employed in this study, reconstructs 3D geometry and color information from 2D images or videos captured from multiple viewpoints. While this allows for the reconstruction of relative object shapes and proportions, it does not directly yield precise measurements, unlike laser-based 3D scanning methods. Despite this limitation, 3DGS offers a practical advantage in forensic applications, as it enables scene reconstruction using multi-angle photographs or videos captured using general-purpose cameras or mobile phones, eliminating the need for specialized equipment. Furthermore, to validate previous claims about the high resolution of 3DGS reconstructions, this study evaluates the dimensional accuracy of 3D-reconstructed scenes generated using 3DGS.

When reconstructing a space using 3DGS, the entire scene is captured holistically, and the resulting output is generated as a unified spatial entity rather than distinct, segmented objects. As previously discussed, this reconstructed 3D scene inherently possesses a relative scale rather than absolute metrics. To align the dimensions of the virtual scene with the physical world, this study implements a method termed ‘Reference Object-based Scale Calibration’.

Specifically, a reference object within the scene is selected and physically measured (Value A in Figure 1). This ground truth measurement is then used to align the virtual dimension of the corresponding object (Value X), thereby applying a global scale adjustment (Value Y) to the entire 3DGS model to establish absolute dimensions. Once this calibration is complete, the dimensions of other target objects—those not physically measured—within the same virtual space (Value Z) are expected to approximate their real-world sizes. Therefore, this study aims to quantitatively verify the accuracy of these inferred measurements (Value Z) within the calibrated virtual environment.

As illustrated in Figure 2, the experimental procedure was conducted in three sequential phases: Phase 1 involved data acquisition and physical measurement of objects in the real-world environment; Phase 2 entailed measuring objects in the virtual space following 3DGS reconstruction; and Phase 3 focused on the quantitative accuracy analysis of the measurement results.

Phase 1 involved Data Acquisition and Physical Measurement. The procedure began with the establishment of a mock crime scene, followed by the capture of photographs and video data for 3DGS reconstruction and the physical measurement of the width, length, and height of objects within the scene. For successful 3DGS reconstruction, images must be captured from various angles while moving along a trajectory to allow the software to detect positional differences between frames; this dynamic methodology contrasts with conventional static forensic photography. Following the data capture process, a specific item was designated as the ‘Reference Object,’ and its exact dimensions were directly measured. Additionally, several other objects of varying sizes were measured for comparative analysis. To complement these physical measurements, the same space was also captured using a 3D scanner to generate a high-fidelity virtual model, which served to provide ground truth data for objects that could not be physically measured on-site.

Phase 2 focused on 3DGS Reconstruction and Virtual Measurement. In this phase, the captured images and videos were processed to digitally reconstruct the scene. The scale of the entire scene was then calibrated by aligning the virtual dimensions of the reference object with its physical measurements obtained in Phase 1. Once this alignment was achieved, the dimensions and proportions of all other objects within the scene were adjusted accordingly to approximate their actual physical sizes. As noted in previous studies by Rangelov et al. (2024) and Lee (2025), camera parameters can influence reconstruction accuracy; thus, careful axis scaling was performed to address potential deviations. Detailed software procedures and environments for this phase are described in Section 3.3.

Phase 3 constituted the Quantitative Accuracy Analysis. In this phase, the geometric accuracy of the 3DGS reconstruction was evaluated by comparing the physical measurements obtained from the actual scene with the virtual measurements derived from the virtual space. To quantify the accuracy of width, length, and height dimensions, three standard metrics—Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE)—were calculated within the R Studio environment. Furthermore, to enhance analytical precision, objects were categorized into groups based on size, or further stratified by specific characteristics such as ‘Thin objects’ and ‘Beyond-room Area’ objects. For these specific subgroups, in-depth additional analyses were conducted utilizing supplemental metrics, including Median Absolute Error (MdAE) and Normalized RMSE (NRMSE).

To reconstruct the photo and video data acquired from the pilot and main experiments into 3DGS models, Jawset’s Postshot software (v0.60) was utilized. This study aims to ensure practical versatility so that forensic practitioners without specialized knowledge in 3D graphics can immediately adopt this technology in the field. Furthermore, to maintain the security of criminal investigations, the software must operate solely using the GPU of a standalone PC, rather than relying on a cloud environment. As Postshot is a commercial software that performs training exclusively on a local GPU, it satisfies these specific requirements.

Regarding the training configuration, all parameters were set to their default values for the pilot study. In contrast, for the main experiment, which involved large-scale datasets, specific adjustments were made to optimize reconstruction quality. Specifically, the ‘Radiance Field Profile’ was set to ‘Splat3’, and the ‘Downsample images’ option was unchecked (disabled) to preserve the original resolution. The ‘Max image count’ was set to match the original quantity of the dataset, with a limit of 1,000 frames for video data, while all other parameters remained at their default settings.

The training duration was approximately 8 min for the pilot study data. For the main experiment, it took approximately 47 min for the photo-based model and 4 h and 14 min for the video-based model.

The computing environments for each experiment were configured as follows: The pilot experiment was conducted on a workstation running Windows 11 Pro, equipped with an Intel Core i9-14900K (3.20 GHz) CPU, 128GB RAM, and an Nvidia RTX 4090 GPU. The main experiment utilized a workstation running Windows 11 Pro, equipped with an AMD Ryzen Threadripper 3,970× 32-Core Processor (3.70 GHz) CPU, 128GB RAM, and an Nvidia RTX 4090 GPU.

Upon completion of training, the 3DGS scenes were extracted in PLY format using the ‘Export Scene’ function, as shown in Figure 6a. Since the raw 3DGS output often contains background noise or floating artifacts, a post-processing step was necessary to obtain a clean virtual scene. To this end, the extracted PLY data were imported into the web-based editor SuperSplat,¹ where unnecessary noise Gaussians were removed and the scene was refined before being saved as a Splat PLY file. While this file shares the same extension as a standard point cloud PLY, it additionally contains the specific attribute information required for Gaussian Splatting rendering. The final scene, refined via SuperSplat, was saved as a Splat PLY file sized 440,009 KB, consisting of a total of 1,909,182 splats.

Subsequently, the Splat PLY data was imported into 3ds Max, where the system unit was configured to millimeters, as a V-Ray 7 Gaussian Splatting geometry object, as shown in Figure 6b. The scale and proportions of the entire geometry object were then calibrated to align precisely with the physically measured dimensions of the reference object. As 3ds Max does not natively support 3DGS, the objects were imported and rendered using the V-Ray plugin.

In the pilot study, the scale of the entire 3DGS scene data was calibrated based on the physical dimensions of the reference object (delivery box). The applied scaling factors were set to 1,984.5% for the X-axis, 1,990% for the Y-axis, and 1,985% for the Z-axis. Subsequently, as illustrated in Figures 7a,b, the dimensions of the props and furniture within the virtual scene were measured by creating and aligning virtual proxy objects to match the boundaries of each target item.

These measurements revealed that, after importing the 3DGS data into 3ds MAX and scaling the data using the reference object, the X-, Y-, and Z-axes were not scaled uniformly, with minor deviations of approximately 0.25–0.27%. This discrepancy was consistent with initial predictions during the experimental design phase and affirmed the necessity of proportion-based adjustments using a reference object rather than uniform size scaling.

For the main experiment, the reconstruction process followed the same workflow established in the pilot study. Consequently, the final reconstructed output was generated as a PLY data file sized 686,091 KB, consisting of a total of 2,976,928 splats.

The TV stand was designated as the reference object. Based on its physical dimensions, the scale of the entire scene geometry was adjusted to 5,180% along the X-axis, 5,076.4% along the Y-axis, and 5,066.2% along the Z-axis. To measure the width, length, and height of the selected target objects within the reconstructed virtual scene, virtual proxy objects were created and aligned with the targets, as illustrated in Figures 8a,b, and their dimensions were subsequently verified.

Upon visual inspection of the reconstructed results, it was observed that moderately sized objects were rendered with clarity, whereas very thin objects, such as paper markers, exhibited indistinct boundaries due to the inherent characteristics of 3DGS Gaussian kernels. Additionally, Gaussian scattering artifacts were identified on reflective surfaces such as desk drawers. Since this study focuses on geometric accuracy rather than visual rendering quality, measurements were conducted by aligning tools with the center of the boundaries, regardless of edge distinctness.

According to the quantitative analysis presented in Table 3, the proposed methodology achieved a high level of geometric accuracy suitable for practical applications. Specifically, the Mean Absolute Error (MAE) for the planar dimensions of Width and Length was measured at 0.25 mm and 1.25 mm, respectively, with Root Mean Square Error (RMSE) also maintaining low values. These quantitative results demonstrate that the reference object-based scale calibration technique effectively synchronizes the overall dimensional proportions of the virtual space with the actual scene.

However, a distinct pattern was observed in the Height dimension. Although the absolute precision was good, with an MAE of 0.65 mm (remaining below 1 mm), the relative metric, MAPE, showed an unusually high value of 200.15%. This discrepancy is interpreted as relative error inflation caused by the minimal physical thickness of objects like A4 paper; despite minute absolute errors, the percentage error is amplified by the extremely small denominator (actual thickness). These pilot study data suggest that while 3DGS accurately specifies the physical location of thin objects, relative proportions may appear distorted. Consequently, these observations served as the rationale for establishing a separate ‘Thin Group’ and adopting supplemental error analysis in the subsequent main experiment.