The existing reconstruction methods for stereoradiography require identifying specific anatomical landmarks projected from the radiographs such that they are not adapted to bones with continuous shapes. To release this limitation, we propose a generic-model-based approach to reconstructing the three-dimensional knee joint from two radiographs in an un-calibrated environment. The generic knee-joint model is built from CT image sets by using the marching cube algorithm first. The control points of the ASM (Active Shape Model) defined on the radiographs are registered to the 3D generic model by using the ICP (Iterative Closest Point) algorithm. The generic model is deformed to obtain the real 3D knee joint by utilizing the RBF (Radial Basis Function). The experimental results show our method can reconstruct the knee joint model accurately and efficiently.