Abstract: Focusing on shape from calibrated images under complex illumination, a two-step reconstruction policy is presented, which combines bounding edge model and level set method. Bounding edge model is used to sample visual hull, while controlling noise by projection color variance threshold. Based on level set method, evolution process of partial differential equation can shrink to target surface from the bounding edge samples, while controlling filter effect by a revisory coefficient in energy model. A technique called varying time steps TVD Runge-Kutta is proposed to reinitialize level set function, which guarantees upwind design of each Euler step. The experiment results show that the deformable initial surface can automatically shrink to model of object from the samples with accuracy and robustness under complex illumination that high light and shadow exist synchronously.