This report addresses the problem of estimating camera parameters from images where object silhouettes only are known. Several modeling applications make use of silhouettes, and while calibration methods are well known when considering points or lines matched along image sequences, the problem appears to be more difficult when considering silhouettes. However, such primitives encode also information on camera parameters by the fact that their associated viewing cones should present a common intersection in space. In this paper, we investigate the problem both on the theoretical and practical viewpoint. In particular, we clarify why, and how, a set of image silhouettes of the same scene give constraints on camera parameters; these constraints justifying a calibration approach. The main contributions of this paper with respect to existing works is first to establish the optimal criterion that camera parameters should satisfy with respect to silhouettes, and second to provide a practical approach based on this criterion. Results on both synthetic and real data are shown to give insights into the method potential for camera calibration.