High-Level Patch-Edge-Corner Based Building Abstraction (Preview)

---

Abstract

Automatic dense 3D reconstruction from images has long been a challenge due to poorly textured or specular surfaces, commonly resulting in noisy or even distorted models.
In this paper, we present an automatic building abstraction method which significantly compresses the architecture whereas maintains high quality. We automatically extract each single building and adopt the planarity assumption for buildings, excluding the clutter brought by the non-planar regions. By employing multi-view consistent images along with the initial stereo, we abstract and refine an architecture's noisy reconstructed model via a Patch-Edge-Curve(PEC) network to implement a global optimal refinement. As our method links the PEC network across multiple overlapping images, we ensures a 3D reconstruction consistent with the real one. Experimental results demonstrates that our method significantly succeeds in recovering piecewise planar surfaces in diverse challenging scenes of distinctive cities.

---

Video