Efficient View-Dependent Sampling of Visual Hulls

In this paper we present an efficient algorithm for sampling visual hulls. Our algorithm computes exact points and normals on the surface of visual hull instead of a more traditional volumetric representation. The main feature that distinguishes our algorithm from previous ones is that it allows for sampling along arbitrary viewing rays with no loss of efficiency. Using this property, we adaptively sample visual hulls to minimize the number of samples needed to attain a given fidelity. In our experiments, the number of samples can typically be reduced by an order of magnitude, resulting in a corresponding performance increase over previous algorithms.