A Robust and Accurate Method for Visual Hull Computation

A novel method for computing visual hull mesh from a sequence of silhouettes is addressed in this paper. Firstly we use genetic algorithm to estimate the bounding box of the scene object from the silhouettes. Secondly, we develop a method for computing the octree of visual hull through a new projection test strategy to determine whether a voxel locating outside, on or inside the visual hull. The projection test depends on a fact that the projection of a voxel in an image is the same as the convex hull of the projections of the cube's 8 vertices. Finally, the visual hull can be extracted from the Octree. In order to get smooth visual hull mesh, we compute the 2D distance to the silhouette, and use it to evaluate the 3D distance to the real visual hull surface. Experiments with several real data sets are presented to validate our algorithm.