Given a set of images captured with a fixed camera while a point light source moves around an object, we can estimate the shape, reflectance property and texture of the object, as well as the positions of the light source. Our formulation is a large-scale nonlinear optimization that allows us to adjust the parameters so that the images synthesized from all of the parameters optimally fit the input images. This type of optimization, which is a variation of the bundle adjustment for structure and motion reconstruction, is often employed to refine a carefully constructed initial estimation. However, the initialization task often requires a great deal of labor, several special devices, or both. In the present paper, we describe (i) an easy method of initialization that does not require any special devices or a precise calibration and (ii) an efficient algorithm for the optimization. The efficiency of the optimization method enables us to use a simple initialization. For a set of synthesized images, the proposed method decreases the residual to zero. In addition, we show that various real objects, including toy models and human faces, can be successfully recovered.