This paper presents an implemented vision-based strategy for computing three-finger stable grasps on unknown planar objects. Its only input is an image of a real unknown curved object instead of synthetic polygonal models. Grasp regions are identified for achieving tolerance around contact points so that errors in finger positioning as well as a finite finger size are considered. It can find expanding and squeezing grasps involving internal and external regions, including holes. Theoretical constraints for force-closure are met while assuring robustness and real-time performance under real-world conditions.