A visual hull is the maximal shape of an object that is consistent with silhouette images. It approximates the shape of a real scene with the limited number of views. Even though the visual hull does not describe perfect shape of an object, the visual hull is frequently used in many computer vision applications because of its simplicity.

In this work, we propose a fast hardware-accelerated 3D modeling method for a smooth visual hull reconstruction. The proposed method consists of two stages. In initial model reconstruction stage, we reconstruct a voxel model from a set of silhouettes and extract an iso-surface as a triangular mesh. We carry out the fast voxel occupancy detection by transforming it to rendering process in GPU. In mesh deformation stage, we iteratively deform the initial mesh model. Through the iterative deformation, the initial mesh becomes to be consistent with input silhouette images. Two constraints, silhouette and smoothness keeps the deformed surface to be smooth, but consistent with the input silhouettes. Thus, our method generate the visual hull that has smooth surfaces while keeping the shape features of the exact visual hull.