This paper presents a series of unique methods to generate a complete B-Reps type solid model, based on the information of stereo image pairs from multiple views that are obtained by taking pictures of polyhedral objects with stereo-cameras. First, we present a method to generate 2D B-Reps that become a basis of topological information of 3D B-Reps, and a hierarchical method to determine corresponding point pairs. Then, by introducing an open shell, a set of visible faces with a boundary, we present a method to generate the open shell regularly in terms of geometry to represent a result of recognition for a single view. Finally, after discussing the problems when unifying plural open shells obtained for multiple views into a solid, we propose a method to overcome them and generate a complete solid (a closed shell) model. This paper asserts that by outputting the B-Reps in terms of the data structure of a 3D CAD a variety of applications become possible, and that, this model, which generates open shells in an intermediate stage, is also effective in reducing shape errors (relative positional errors).