From Multiple Stereo Views to Multiple 3-D Surfaces

• Authors: P. Fua
• Venue: International Journal of Computer Vision
• Year: 1997

We present a framework for 3-dimensional surface reconstruction thatcan be used to model fully 3-D scenes from an arbitrary number ofstereo views taken from vastly different viewpoints. This is a keystep toward producing 3-D world-descriptions of complex scenes usingstereo and is a very challenging problem: real-world scenes tend tocontain many 3-D objects, they do not usually conform to the 2-1/2-Dassumption made by traditional algorithms, and one cannot take it forgranted that the computed 3-D points can easily be clustered intoseparate groups. Furthermore, stereo data is often incomplete andsometimes erroneous, which makes the problem even more difficult.By combining a particle-based representation, robust fitting, andoptimization of an image-based objective function, we have been able toreconstruct surfaces without any a priori knowledge of their topology and in spite of the noisiness of the stereo data.Our current implementation goes through three steps—initializing a set ofparticles from the input 3-D data, optimizing their location, and finallygrouping them into global surfaces. Using several complex scenescontaining multiple objects, we demonstrate the competence of this methodand its ability to merge information and thus to go beyond what can be donewith conventional stereo alone.