Multi-Level 3D Reconstruction with Visibility Constraints

In this paper, we examine the problem of 3D reconstruction from a large number of images annotated with camera pose information. Our proposed method involves computing 3D descriptions of the scene at multiple levels of representation (vertex, line segment, and surface) simultaneously. A major problem encountered during bottom-up processing is the uncertainty of low-level feature correspondence, which often results in mis-matches. We observe that visibility constraints from surface-level information can help in resolving ambiguity. This motivates our design of a hypothesis-based 3D correspondence system that integrates information from multiple levels. Experimental results on reconstructing urban scenes are presented. 1 Introduction Large-scale 3D reconstruction is a challenging problem in computer vision. In this paper, we assume our input images are controlled, i.e. annotated with information on each camera's orientation and position. This allows us to perform reconstruction in a spa..