Indoor Structure-aware 3D Gaussian Splatting for Panorama-Style Motion

Abstract

Recent advances in 3D Gaussian Splatting (3DGS) have enabled real-time novel view synthesis (NVS) with impressive quality in indoor scenes. However, high-fidelity rendering typically requires densely captured images covering the entire scene, which limits accessibility for casual or mobile users. This study aims to develop a practical 3DGS-based framework suitable for panoramic, rotation-dominant motion captured by a single mobile camera in indoor environments. Such motion patterns inherently produce narrow baselines and unreliable pose estimations, especially in textureless indoor scenes, lead- ing to inaccurate geometry reconstruction such as COLMAP. To address these challenges, the proposed framework leverages rough geometric priors, including mobile device poses and monocular depth esti- mation, while exploiting the planar structures frequently observed in indoor environments. A plane scaffold assembly method is introduced to generate geometrically consistent 3D points along planar surfaces. During 3DGS optimization, a stable pruning strategy is applied to improve geometric alignment in textureless regions. Geometric and photometric corrections are employed to mitigate in- consistencies caused by motion drift and auto-exposure variations inherent to mobile devices. Experiments on both real and synthetic indoor datasets demonstrate that the proposed approach achieves photorealistic rendering quality, outperforming state-of-the-art methods under minimal-effort, panorama-style data acquisition. These results highlight the potential for practical panoramic view syn- thesis and robust object placement in indoor scenes.