One of the difficulties in light-probe 3D vision measurement is the stereo matching of the light-spot image points in the left and right images. Therefore, the SoftPoist algorithm was adopted to transform the difficulty to a position and orientation estimation problem under the condition that the correspondences of the 2D image points and the 3D object points are unknown. By combining the correspondence problem and the position and orientation estimation problem together to perform bundle optimization, the matching of the light-spot image points and the spatial light-emitting-spots of the light-probe could be completed, respectively, in order to achieve the stereo matching of the light-spot image points in the left and right images. The proposed method is based on model and is not related to the spatial point distribution, point number and the image gray level. The experiment results show that the proposed method is feasible and can be also used for feature points recognition.