Over the past years, Structure-from-Motion calibration algorithms have become widely popular for many applications in computer graphics. From an unordered set of photographs, they manage to robustly estimate intrinsic and extrinsic camera parameters for each image. One major drawback is the quadratic computation time of existing algorithms. This paper presents different strategies to overcome this problem by only working on subsets of images and merging the results. A quantitative comparison of these strategies reveals the trade-off between accuracy and computation time.