This paper presents a Bayesian framework for Visual Hull reconstruction from multiple camera views with a 3D sampling scheme based on an irregular 3D grid, which becomes regular once projected onto the available views. The probabilistic framework consists in establishing a foreground probability for each pixel in each view rather than segmenting in order to obtain binary silhouettes of the foreground elements. Next, a Bayesian consistency test labels the occupancy of each image-adapted 3D sample. The proposed method, using image-adapted 3D sampling in the Bayesian framework, is compared to a shape-from-silhouette implementation with image-adapted voxelization, where the input data are binary silhouettes instead of probability maps; we also compare its performance to a state-of-the-art method based on regular 3D sampling with binary silhouettes and SPOT projection test.