This paper presents a 3-D water-flow reconstruction algorithm for use on stereo-camera systems. It contains two major contributions. The first is the stereo-camera calibration method. A vanishing line is used to get a closed-form solution of the parameters of one camera and then the bicamera parameters are recursively updated for more precise results. The second contribution is the stereo matching of particles in the water flow. The difficulty here is that the correspondence of the maximum-correlation-coefficient method is not right for all water flow particles, because the illumination environment is not uniform and the aperture of the two cameras is different. We use relaxation labeling and template matching to cluster the water particles into group sets and compute the best correspondence of these sets based on epipole and camera geometry. The experimental results are reliable.