Active Object Modeling with VIRTUE

Abstract. This paper presents a vision system for the task of actively acquiring and modeling the geometry of an unknown object. Using an active trinocular stereo head (VIRTUE), sensed 3-D line segments are grouped into a polyhedral volumetric model through the aid of a constrained Delaunay triangulation. Partial models and a viewpoint enumeration scheme are used to guide the image acquisition process and to determine `where to look next'. Results of the active vision recovery of a number of objects are provided with their associated volumetric and surface errors.