A semi-automatic system to infer complex 3-D shapes from photographs

摘要

We present a semi-automatic image-based modeling system to make 3-D models from photographs. The human operator intervenes only for simple and straightforward tasks. The design of such a system is motivated by applications where the geometric shapes of the objects are complex, and the requirement on the appearance of the models is high. Ease of use is another consideration. We use binocular stereo as the basic model recovery tool, and propose a new volumetric matching algorithm that fully explores the spatial coherence between the two images. Inaccurate matches are mostly clustered around the border area between the foreground and the background thus easy to be manually removed. To make the system more practical, we insist on using uncalibrated cameras. Experimental results on different types of real objects are presented.