Extraction of Concise and Realistic 3-D Models from Real Data

We have developed a novel algorithm for extracting concise 3-D surface models of a scene from real sensor data that can be used for virtual reality modeling of the world. Our algorithm produces an accurate and realistic description of a scene even in the presence of sensor noise, sparse data and incomplete scene descrip-tions. We have demonstrated order of magnitude reduction in the number of faces used to describe a scene using data from multibaseline omnidirectional stereo, structure from motion, and a light stripe range finder. The algorithm takes as input a surface mesh describing the scene and a co-registered image. It outputs a surface mesh with fewer vertices and faces to which the appearance of the scene is mapped. Our simplification and noise reduction algorithm is based on a segmentation of the input surface mesh into surface patches using a least squares fitting of planes. Sim-plification is achieved by extracting, approximating and triangulating the bound-aries between surface patches