Volumetric visual hulls have become very popular in many computer vision applications including human body pose estimation and virtualized reality. In these applications, the visual hull is used to approximate the 3D geometry of an object. Existing volumetric visual hull construction techniques, however, produce a 3-color volume data that merely serves as a bounding volume. In other words it lacks an accurate surface representation. Polygonization can produce satisfactory results only at high resolutions. In this study we extend the binary visual hull to an implicit surface in order to capture the geometry of the visual hull itself. In particular, we introduce an octree-based visual hull specific adaptive sampling algorithm to obtain a volumetric representation that provides accuracy proportional to the level of detail. Moreover, we propose a method to process the resulting octree to extract a crack-free polygonal visual hull surface. Experimental results illustrate the performance of the algorithm.