Methodological workflow for creating photorealistic neoclassical architecture datasets

Abstract

Computer vision models for architectural recognition and digital heritage require high-quality, labelled, real-world data, which can be obtained through various 3D reconstruction techniques. Alternatively, synthetic digital versions can be artist-driven, utilizing computer graphics and 3D modeling techniques. In this paper, we propose a methodological workflow for creating the ArcherVerse dataset, tailored for modelling neoclassical architecture. We specifically focus on re-creating De La Salle University (DLSU), which is a university that contains neoclassical structures. Our proposed approach integrates high-poly manual modeling, physically based rendering (PBR), block-out, and scene compositing to reconstruct the ornate facades and intricate motifs characteristic of the neoclassical style. A user evaluation involving 16 participants possessing expertise in 3D rendering yielded mean photorealism and artistic quality scores exceeding 4.0 out of 5.0. The results demonstrate that artist-driven manual modelling can provide “clean” geometric ground truth data as an alternative to traditional 3D scanning methods for reconstructing architectural buildings. The ArcherVerse dataset can serve as a benchmark for sim-to-real research in CV applied to neoclassical architecture scene understanding.