A novel approach for computing exact visual hull from silhouettes

Modeling from silhouettes is one of the main technologies for 3D reconstruction in computer vision. The existing methods are generally divided into two classes: volume-based approaches and surface-based approaches. The former are robust for complex objects but inaccurate, while the latter are accurate but time consuming and usually suffer from numerical instabilities. This paper proposes a novel method to compute exact visual hull by taking advantage of surface-based approaches while retaining the simplicity and robustness of volume-based approaches. To this aim, a line-based model is built for each cube on the edge of octree volume structure of object. Convex surface points of visual hull in line-based model are searched by the idea of surface-based approaches. Experiments with real data are conducted to test the efficiency and exactness of our method.