Surface Reconstruction
We recover the 3D shape of an object from its silhouettes. The main innovation is that no correspondences are computed, significantly reducing computational complexity. The image at right shows the reconstructed surface of a pear.
Background & Objective: 3D-from-silhouettes has long been studied in computer vision as a cheap way to 3D-scan objects. Current methods have problematic time complexity and generally produce very jagged 3D surfaces.
Technical Discussion: Technically, this is known as the convex hull problem. Previous methods requires expensive point matching to determine where silhouettes cross in 3-space. We project the problem into a dual space and show that by exploiting simple continuity principles in that space, a smooth manifold approximating the convex hull can be estimated without matching.
Publications:
Technology Area: Computer Vision
Modification Date: September 12, 2007