Reconstructing complex surfaces from multiple stereo views
P. Fua
SRI International
333 Ravenswood Avenue, Menlo Park, CA 94025, USA
We reconstruct surfaces or arbitrary topology from multiple stereo views. This is a difficult problem because the data typically is noisy and the 2-1/2--D assumption of traditional stereo algorithms does not hold.
To overcome these difficulties, we first fit local surfaces to the stereo data. We then optimize their positions by treating them as particles that interact with each other through forces that tend to align them and by minimizing an image-based objective function. The first step eliminates the isolated errors and reduces the large number of 3--D points to a smaller and more manageable set of particles. The second allows for more precise modeling and more effective grouping.
We will demonstrate our technique on several indoor and outdoor scenes and produce 3D--models of complex surfaces such as a complete head.