Flexible Planes Extraction for Interactive-Friendly Multiview Stereo in Architectural Scenes

This paper proposes an automatic planes extraction approach to recover hypothesis planes of architectural scenes, being easily adopted in interactive-friendly multiview reconstruction systems. There exist many interactive scene reconstruction algorithms designed to model these real worlds, but they are very labor-intensive and difficult to verify whether the resulting 3D models are accurate or not. Motivated by the strongly piecewise planar property of architectural scenes, the hypothesis 3D planes can be generated automatically and flexibly in our work, being faster, more accurate and labor-saving than that of manually processing. Firstly, initial oriented points are reconstructed by using structure-from-motion and patch-based multiview stereo. Secondly, non-parametric density-based clustering technique is exploited to extract plane normal vectors without any presumption on the prior distributions of the normals of the planes, unlike famous Manhattan-world assumption. Then, the offsets of the planes can be estimated to specify the locations of the assumed 3D planes. Each group of the normal vectors and offset distances along the corresponding normals can determine certain candidates of plane structures. The obtained hypothesis planes are very beneficial to interactive-friendly multiview reconstruction with its efficiency and labor-saving, without measuring the sites, locating and digitizing architectural planes or converting existing CAD data.