Résumé : The 3D visibility skeleton is a data structure used to encode global visibility information about a set of objects. Previous theoretical results have shown that for $k$ convex polytopes with $n$ edges in total, the worst case size complexity of this data structure is $\Theta(n^2 k^2) $ [Brˆnnimann et al. 07]; whereas for $k$ uniformly distributed unit spheres, the expected size is $\Theta(k)$ [Devillers et al. 03]. In this paper, we study the size of the visibility skeleton experimentally. Our results indicate that the size of the 3D visibility skeleton, in our setting, is $ C\,k\sqrt{n\,k}$, where $C$ varies with the scene density but remains small. % This is the first experimentally determined asymptotic estimate of the size of the 3D visibility skeleton for reasonably large $n$ and expressed in terms of both $n$ and $k$. We suggest theoretical explanations for the experimental results we obtained. Our experiments also indicate that the running time of our implementation is $O(n^{3/2} k\log k)$, while its worst-case running time complexity is $O(n^2k^2 \log k)$.