Accuracy of three-dimensional Gaussian Splatting for virtual crime scene reconstruction

With the rapid advancement of artificial intelligence, three-dimensional (3D) Gaussian Splatting (3DGS), which reconstructs 3D data from standard photographs and videos, has garnered increasing attention in digital forensic applications. This study evaluated the quantitative accuracy of 3DGS-based virtual crime scene reconstruction to determine its suitability for forensic documentation. To this end, a mock crime scene was constructed, and both photographs and videos were captured using a DSLR camera to generate a virtual environment through 3DGS. Since the generated environment inherently possesses only relative scale, a ‘Reference Object-based Scale Calibration’ method was employed to establish absolute dimensions by adjusting the scale of the entire virtual space based on the physical measurements of a single reference object. The reconstructed object dimensions were then compared with actual measurements in two phases: a preliminary test involving seven objects and a main test involving 13 objects provided by the Seoul Metropolitan Police Agency. The results demonstrated millimeter-level accuracy, with mean measurement errors ranging from 0.25 to 0.65 mm in the preliminary test and from 1.73 to 3.58 mm in the main test. Notably, while larger objects such as desks and doors exhibited stable reconstruction accuracy, smaller or thinner items like bloodstains showed higher relative errors due to scale-induced artifacts; however, their absolute physical precision remained intact. Overall, these findings underscore the potential of 3DGS as a reliable and practical tool for the digital preservation and reconstruction of crime scenes.