Towards Unsupervised Reconstruction of Architectural Models

摘要

Architectural blueprints offer a concise, clear and high-level description of the structure of a building. On the other hand, state of the art reconstruction pipelines can nowadays produce dense point clouds or high-polygon meshes without any human intervention from a set of digital images or video. We present a fully automated structure and motion framework capable of capturing the expressivity of a high-level description without sacrificing the minute details of an accurate reconstruction. Our resulting architectural models, composed of textured high-level geometric primitives, capture the overall structure of a building and give birth to a more tractable and abstract model of the imaged scene, thereby narrowing the semantic gap in 3D reconstruction. Several examples display our system in action.