Video-driven Gaussian splatting for as-built building geometry with energy simulation

This study introduces Gaussian-based envelope modeling automation (GEMA), a framework that uses Gaussian splatting to convert video footage into closed-surface 3D models suitable for building energy modeling (BEM). Focusing on exterior envelope geometry (LoD3), GEMA reconstructs the planar building surfaces needed for BEM. The pipeline integrates structure-from-motion, 2D Gaussian splatting, α-blending, and planar-SSIM to enhance surface accuracy. Tested residential buildings, GEMA achieved geometric deviations under ≤ 20.47%, and ASHRAE suggested hourly CV(RMSE) ≤ 30% for simulations compared to ground-truth simulations. Densification improved surface fidelity and simulation stability, with a resolution sensitivity analysis showing that annual predictions remained stable within ∼0.5–12.5% of a ground-truth baseline. A planar similarity score (P-SSIM) based on SSIM and a discrete adaptation of the 3D Radon transform is proposed to evaluate simple meshes. GEMA supports construction digital workflows by enabling automated reconstruction of as-built building envelopes from video data, producing simulation-ready geometry for energy analysis and sustainable building design. • GEMA converts drone video to BEM-ready LoD3 meshes via 2D Gaussian splatting. • P-SSIM metric evaluates mesh planarity using SSIM and 3D Radon transform. • Densification improves surface fidelity and energy prediction stability. • Scalable video-driven workflow reduces reliance on LiDAR or manual modeling.