A Two-Stage Diffusion Framework for 3D Building Reconstruction from Single Nadir Images
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Reconstructing 3D urban geometry from single nadir aerial images is an ill-posed problem due to severe vertical ambiguity and occluded building facades. To resolve this, this thesis proposes GeoPriorPC, a novel two-stage generative framework bridging foundational 2D image priors and unstructured 3D point clouds. Stage 1 fine-tunes a Latent Diffusion Model (LDM) to extract a compressed geometric prior, specifically, oblique surface normals, from flat nadir inputs. Stage 2 employs a Diffusion Point Transformer (DiPT) backbone. It utilizes a dual-conditioning mechanism, integrating global adaptive Layer Normalization (adaLN-Zero) and local spatial cross-attention, to structurally guide the 3D denoising process using the learned 2D prior. Evaluated on the AIM2PC benchmark, GeoPriorPC establishes a new state-of-the-art, achieving a 44% relative improvement in F1-score and reducing Earth Mover's Distance by 49%. By explicitly resolving depth ambiguity, this framework successfully synthesizes the accurate architectural manifolds critical for urban digital twins.