Human face is a 3D object with shape and surface texture. 3D Morphable Model (3DMM) is a powerful tool for reconstructing the 3D face from a single 2D face image. In the shape fitting process, 3DMM estimates the correspondence between 2D and 3D landmarks. Most traditional 3DMM fitting methods fail to reconstruct an accurate model because face shape fitting is a difficult non-linear optimization problem. In this paper we show that landmark weighting is instrumental to improve the accuracy of shape reconstruction and propose a novel 3D Morphable Model Fitting method. Different from previous works that treat all landmarks equally, we take into consideration the estimated errors for each pair of 2D and 3D corresponding landmarks. The landmark points are weighted in the optimization cost function based on these errors. Obviously, these landmarks have different semantics because they locate on different facial components. In the context of the solution of fitting is approximated, there are deviations in landmarks matching. However, these landmarks with different semantics have different effects on reconstructing 3D faces. Thus, it is necessary to consider each landmark individually. To our knowledge, we are the first to analyze each feature point for 3D face reconstruction by 3DMM. The weight is adaptive with the estimation residuals of landmarks. Experimental results show that the proposed method significantly reduces the reconstruction error and improves the authenticity of the 3D model expression.