3D recovery of object shape and camera motion from 2D image sequence is formulated as a non-linear optimization problem. Such a non-linear approach ensures the most precise solution, but it involves computational difficulties, i.e. complexity and instability. This paper describes a practical procedure to solve the non-linear optimization problem for object shape and camera motion recovery. First, we discuss the imaging model for large objects such as buildings; one shot and pseudo wide angle models, where the camera motion is modeled as a uniform motion along circular path with gazing at the center. Then, the generalized initial values for circular motion is proposed and "mean appearance ratio (MAR)" of feature points is defined as a measure of stability. Experimental results with synthetic images show that 3D recovery was stably performed from image sequences with MAR over 15%, quasi-stable with over 8.5% for one shot model, and stable with over 10.3%, quasi-stable with over 4. 4% for pseudo wide angle model. 3D shape and motion using real images were recovered successfully from MAR 17% for one shot model and 6% for pseudo wide angle one. These results have demonstrated the robustness of the generalized initial values and effectiveness of MAR for large object shape from uncalibrated camera motion problem.