Abstract: Fundamental matrix estimation is a key problem of computer stereo vision, and its estimated accuracy will have a direct influence on corresponding points match and the reconfiguration calculation of feature point’s coordinate. In this paper, studies on the present fundamental matrix estimation methods are carried out,and their performance, merits and demerits are analyzed. Based on these work, an improved Torr M-estimators is put forward. By selecting first value（F₀）based on competition and using the mask printed with standard pattern to assistant computation, outliers are wiped off effectively, the accuracy of fundamental matrix estimation is improved, and the epipolar geometric relation is resumed precisely.