Floorplan-aware camera pose estimation for automated 3D reconstruction of indoor environments from panoramic images

초록

Image-based 3D reconstruction is an important technology that produces as-built 3D layouts, enabling owners and site managers to confirm defects and deviations to verify as-built condition. However, most reconstruction pipelines still rely on trained models and manual processing, which limits practical deployment because they require specialized expertise. This paper introduces a floorplan-based algorithm that automatically estimates a 3D layout from a single panorama without model training. The workflow aligns the input panorama to a floorplan-based virtual model by maximizing pixel-wise similarity between the input and synthetic panoramas and employing hierarchical camera pose estimation. The method is evaluated on three datasets covering seven distinct layout types. Across these datasets, the method achieves relative improvements of 45.3 % in accuracy, 43.6 % in consistency, and 26.1 % in adaptability over the best-performing baseline. This method supports nonexperts in documenting as-built conditions for digital twins without expert manual tasks or learning-based processing.

키워드