Linear octree by volume intersection using perspective silhouettes

Abstract

Abstract Recovery of 3D shape from the silhouettes of the object has gained interest among the researchers in computer vision. Volume intersection is the underlying paradigm wherein the object is reconstructed by intersecting the volumes generated by sweeping the silhouettes along the respective viewing directions. Several attempts have been made to design efficient algorithms for shape recovery using this paradigm. These algorithms differ from one another in the object representation schemes, perspective versus orthographic projection and the viewing directions. An efficient algorithm has been proposed earlier to generate a linear octree of the object from its orthographic silhouettes along three mutually orthogonal views. Generally, there is some degree of perspective viewing in any real life situation and orthographic silhouettes are unrealistic. In this paper we propose an algorithm to generate a linear octree from silhouettes in perspective projections. The algorithm makes use of a novel technique of segment tree structure to achieve efficiency. Experimental results are also reported.