Focusing on the complete reconstruction of non-convex surface under complex illumination, a multi-view reconstruction method is presented. Stereo matching is translated into an issue of color variance minimizing. A threshold is used to balance the contradiction between sampling number and sampling noise, and the ambiguity in stereo match is avoided. By using level set method to deduce the equation of curve surface moving, the object surface is approximated by sign distance function, and the sampling noise is filtered by minimizing the elasticity potential energy, and the robustness for occlusion and illumination condition is strengthened. The experimental results show that the deformable initial surface can shrink to model of object with accuracy and robustness under complex illumination, and the reconstructed time is saved 31~39% than the Jin’s algorithm.