Octree refinement using depth information for objects having concavities

Abstract

All techniques for constructing 3-D models from a sequence of views have an inherent problem: if some parts of objects cannot be described by the edge of the projections, these parts of the objects will not be described by reconstructed 3-D models. A method to overcome this problem is presented by refining the octree models using depth images. In this method, an initial octree is constructed from a sequence of silhouettes. Depth images are captured from such viewing angles that the concavity information is sufficient. Each depth image is then converted into a sequence of binary images. Each binary image associates with the corresponding depth value. The octree is progressively refined by each binary image, based on examining the overlay of octree model, binary image, and depth value in 3-D space. >