In this paper, an efficient error analysis of a real-time vision-based pointing system is proposed. We use two cameras to implement the pointing system according to a simplified 3D reconstruction scheme which is based on image feature extraction, homography, and 3D geometry. To that end, we study the relationship between image noises and the ultimate reconstruction errors, and develop efficient methods to find the error range of the latter given a range of the former. Experimental results show that the proposed approach can find the error range satisfactorily. Accordingly, users of similar pointing systems can get more robust pointing results by selecting a special pointer location, or possibly a special pair of cameras, that will result in minimal range of pointing error.