A Mixed-Initiative Procedural Workflow for Fulldome Scene Assembly from Layered 2D Illustrations

Abstract

Transferring layered 2D illustrations to immersive fulldome environments is challenging: manual 3D modeling does not scale to scenes containing several hundred elements, while AI-based monocular depth estimation (MDE) suffers from domain shift on stylized and panoramic content and lacks deterministic artist control. Existing approaches address procedural automation, artist-driven calibration, and dome-specific layout separately rather than in one mixed-initiative pipeline. This study proposes a mixed-initiative procedural workflow that transforms layered 2D illustrations into 3D fulldome scenes with stable placement and billboard orientation under a fixed dome camera, requiring no recomputation under camera rotation. The system uses a horizon-parameterized depth mapping heuristic converting pictorial height into radial depth, cylindrical mapping for 360° layout, and figure-derived terrain fitting via implicit surface reconstruction. Hierarchical parametric calibration preserves artist control over depth and scale at global and category levels. The workflow was tested on a 400-element historical visualization project. It achieved full procedural placement coverage across eligible scene elements with no per-element manual correction, completed in three parameter-adjustment iterations. Across 400 automatically identified overlapping figure pairs, determined by pixel-level alpha-mask intersection without manual selection, occlusion-order consistency with source-layer relations was 357/400 (89.25%). The mean ground-contact correction after raycasting was 0.62 m, with 78% of corrections within ±1 m. Production time was estimated to improve approximately 30-fold. Rule-based depth mapping with mixed-initiative calibration thus offers a viable, deterministic, artist-controllable approach for fulldome scene assembly from stylized panoramic content.