XR 스마트글래스를 이용한 3D Gaussian Splatting 기반 디지털 트윈의 재관측을 통한 모델 정밀화 파이프라인

초록

Digital twins are virtual replicas of physical environments with growing industrial applications; however, their initial construction often suffers from limited rendering fidelity in under-observed directions owing to restricted viewpoints. This study proposes an end-to-end pipeline that improves the model fidelity of 3D Gaussian splatting (3DGS)-based digital twins using a newly observed viewpoint video from XR glasses. The pipeline comprises five stages: keyframe selection, visual localization, refinement necessity detection, 3DGS model refinement, and separate evaluation. Experiments on two indoor environments showed that incremental refinement improved the new-viewpoint PSNR from 19–20 dB to over 30 dB while limiting the existing viewpoint degradation to within 1.36 dB. Full retraining achieved 2.11–2.35 dB higher new-viewpoint quality but showed comparable existing-view preservation, suggesting that the explicit representation of 3DGS inherently limits catastrophic forgetting, regardless of the training strategy.

키워드