Shape-merging and interpolation using class estimation for unseen voxels with a GPU-based efficient implementation

The merging of multiple range images obtained by 3D measurement systems for generating a single polygon mesh, and processing for filling holes caused by unmeasured data or insufficient range images are essential processes for CAD, digital archiving of shapes, and CG rendering. Many of the existing processes that have been proposed for merging and interpolating multiple shapes do not function well when the number of range images is small. In this paper, the space carving method is improved, and an interpolation algorithm is proposed which is capable of producing stable results even when the number of range images is small. In the proposed method, not only the observed voxels in a signed distance field, but also unseen voxels are determined as either inside or outside of an object using Bayes estimation. Characteristics of the proposed method include that closed surfaces are always obtained, and a GPU-based, efficient implementation is possible. In addition, in the case that the range image is obtained using an active stereo method, high precision estimation results can be achieved by using information regarding the light sources.